# scrapity scrap scrap # Scrap?? Perhaps suprisingly, models fared much better at this sort of problem when they were told to ‘[[Think step]] by step’, [leading them to….] This guidance can [[come from]] prompting techniques like Chain-of-Thought (CoT). CoT prompting involves instructing the model – either by providing examples (few-shot) or by using direct commands like 'Let's [[think step]] by step' (zero-shot) – to articulate intermediate steps before providing a [[final answer]]. Instead of generating only a conclusion, the model is steered towards producing a sequence that outlines a potential reasoning process, such as "Step 1: Identify the letters... Step 2: Count the 'r's... Step 3: State the total." This technique works by leveraging the standard mechanisms of autoregressive LLMs. Recall that these models generate text token-by-token, with each prediction conditioned on the preceding context, and use self-attention to weigh the relevance of earlier parts of that context. When CoT prompting is used, the initial instruction biases the model to start generating tokens that represent the first step of a reasoning process. These generated tokens are then added to the ongoing context. For the next prediction, the model attends to this augmented context, including the first step, which influences it to generate tokens representing the second step, and so on. The intermediate steps thus become part of the conditioning context for subsequent steps, effectively creating an explicit 'scratchpad' within the model's context window that guides the generation towards a structured, step-by-step output. Their ability to produce coherent reasoning chains derives from learning complex statistical patterns in their [[training data]], which includes vast amounts of text exhibiting step-by-step explanations, arguments, and problem-solving procedures. CoT prompting activates these learned patterns, causing the model to generate text that mimics, or simulates the form and sequence of reasoning it has observed, without actually understanding the logical or conceptual relationships involved. The quality and reliability of this simulation depend heavily on factors like model scale – CoT effects are much more pronounced in very large models – and refinement techniques like [[Instruction Fine]]-Tuning (IFT) and [[Reinforcement Learning]] from [[Human Feedback]] (RLHF), which improve the model's ability to follow instructions and generate outputs that align with human judgments of coherence and accuracy. Many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) seem to integrate these structured reasoning capabilities more deeply, often producing step-by-step outputs without explicit CoT prompts. Despite being a simulation, this capability is highly effective. Models guided by CoT principles demonstrate significantly improved performance on a wide range of benchmarks requiring multi-step reasoning, such as arithmetic word problems, commonsense inference, and symbolic manipulation tasks. For many problem-solving purposes across various domains, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions where simpler prompting methods fail. This effectiveness connects directly to the potential for enhancing [[philosophical understanding]]. The simulation of reasoning allows LLMs to generate outputs structured like arguments, presenting sequences of claims linked by inferential markers and intermediate steps. As established earlier, it is precisely this type of structured output, rather than isolated pronouncements, that typically facilitates the mapping of [[dependence relations]] central to [[philosophical understanding]]. These simulated arguments can be complex and detailed, offering substantive textual material for analysis. [Placeholder for examples of complex CoT philosophical arguments]. Furthermore, the philosophical utility of these simulated arguments mirrors our engagement with human-authored texts. A philosophical argument need not be perfectly sound, nor its conclusions fully endorsed, for it to be intellectually valuable. Engaging with an argument – identifying its premises, scrutinising its steps, evaluating its inferences, even if flawed – prompts the reader to critically examine the purported dependencies, clarify concepts, and potentially revise or refine their own representation of the relevant conceptual network. The same potential for refinement exists when engaging with the structured output of an LLM. The fact that the reasoning is simulated does not preclude the output from serving as a useful artefact for philosophical reflection, analysis, and the enhancement of the user's own understanding, particularly given the 'epistemically undemanding' nature of understanding adopted in §1. 4. Enhancing Understanding via Chain-of-Thought Outputs The structured outputs generated via CoT prompting can directly enhance [[philosophical understanding]] as defined by the Dellsén framework – the accurate and comprehensive representation of dependence networks. [[The value]] lies not in assuming the LLM 'understands' the dependencies it articulates, but in how the generated text can prompt refinements in the user's representation. • Improving Accuracy: By presenting a step-by-step derivation, CoT outputs allow the user to scrutinise the purported [[dependence relations]] involved. If the LLM generates an argument linking concept A to concept C via an intermediate step B, the user can evaluate the claimed dependence of C on B, and B on A. Identifying a weak or fallacious link in the chain prompts the user to correct their own mental model, potentially removing an inaccurately represented dependency. Conversely, a valid step might reinforce or confirm a dependency the user held tentatively. • Increasing Comprehensiveness: The intermediate steps generated by CoT can introduce new nodes or relations into the user's dependence network. The model might highlight a conceptual distinction, an implicit premise, or a relevant factor that the user had not previously considered, thereby expanding the scope of their representation. It might articulate dependencies between elements that the user had considered disparate. Furthermore, by tracing a line of reasoning, CoT can help delineate negative dependencies – showing, for instance, why a certain conclusion does not follow from a given premise under a specific interpretation, thus refining the boundaries of the represented network. • Clarifying [[Dependence Types]]: Philosophical understanding involves grasping how things hang together. The step-by-step structure can help disambiguate the nature of the proposed dependencies – are they logical, causal, conceptual, constitutive, or mereological? While the LLM might not explicitly label them, the context provided by the intermediate steps can allow the user to better infer the type of relationship being posited, leading to a more precise representation. Crucially, within the Dellsén framework, this enhancement does not require the user to have justification for, or even belief in, every proposition generated by the LLM. The framework is "epistemically undemanding" in this sense. The value derives from the potential of the generated structure to prompt a re-evaluation and refinement of the user's existing representation, leading to a potentially more accurate or comprehensive map of the relevant dependence network, even if parts of the generated chain are ultimately rejected upon reflection. ## Scrap?? - This distinction between the system's internal state and its output's potential function resonates with the analysis by Butlin and Viebahn (forthcoming) regarding AI assertion. - They argue that for an output to be an assertion, it must possess a 'descriptive function' – the function of conveying information to a consumer. - Crucially, they contend that pre-trained LLMs, optimised solely to predict likely word sequences, produce outputs lacking this descriptive function. - Their outputs mimic assertion form, but their function is merely sequence completion, not information conveyance. - However, LLMs fine-tuned specifically for 'groundedness' or 'correctness', like LaMDA or Sparrow, undergo training that selects for accuracy and alignment with known sources. - This fine-tuning process, Butlin and Viebahn suggest, means the outputs of such systems may acquire a descriptive function, serving the purpose of providing trustworthy information, even if the systems themselves fail other conditions for genuine assertion, such as sanctionability. - This highlights that an output's functional properties, like potentially having a descriptive function, can be distinct from whether the generating system possesses understanding . One could argue that the origin of the text, whether from a thinking being or a non-understanding system, matters less than its content and its ability to make the human reader think philosophically. To take this argument further, one might use a concept like 'descriptive functions', possibly drawing on work by figures like Butler [Citation needed if specific work is intended]. The basic idea here is that even if an LLM output is not backed by the usual kind of authorial intention or understanding, the text itself can still function descriptively. It produces outputs designed to be useful, claiming to offer descriptions – perhaps unstated descriptions – of the world or parts of it. These outputs, because of their structure and content, could potentially contain or lead to philosophical insights, regardless of their source. This view, however, faces its own problems. A likely counter-argument might use the analogy of infinite monkeys eventually producing Shakespeare; if LLM outputs are just statistically generated sequences, why should their descriptive functions be taken seriously as potential sources of philosophical insight? This objection highlights the need for more study into why LLM-generated text, seen as having descriptive functions, might be considered important for knowledge or able to contribute meaningfully to philosophical discussion. Dealing with this requires a closer look at how LLMs work and the exact nature of these proposed descriptive functions. Historically, the capacity of LLMs to provide reasoned arguments has been subject to doubt. Earlier models, for instance, often produced outputs where the reasoning presented was weak, inconsistent, or seemingly confabulated [Optional: Add brief examples here]. The development of prompting techniques such as Chain-of-Thought (CoT), however, marked a significant shift in approach. CoT was designed specifically to elicit intermediate reasoning steps from LLMs, thereby improving performance on tasks requiring sequential reasoning processes. Common methods for eliciting CoT include few-shot prompting – providing examples of step-by-step reasoning within the prompt – and zero-shot prompting, which uses triggers like the phrase 'Let's think step by step'. Furthermore, many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) appear to incorporate CoT-like structured reasoning capabilities more implicitly; they often generate step-by-step explanations or arguments without explicit CoT triggers, suggesting this methodology is now somewhat automated or integrated via system prompts or subsequent fine-tuning stages. Technically, CoT leverages the underlying architecture of autoregressive Transformers. As detailed previously, these models predict the next token based upon the preceding sequence, utilising self-attention mechanisms to weigh the importance of different elements within the context. CoT prompting guides this generative process: instead of predicting the final answer directly, the model is steered towards generating tokens which represent intermediate steps in a reasoning sequence. Each generated step is then appended to the context. Consequently, the model's attention mechanism considers this augmented context – including the steps generated thus far – when predicting the subsequent token. This iterative process allows the model to condition later parts of its output on earlier parts, effectively employing the generated text as an explicit 'scratchpad' to manage the simulated reasoning process. It must be acknowledged that this process constitutes a simulation of reasoning, driven by the model learning statistical patterns from vast quantities of text data that include examples of arguments and explanations. LLMs do not manipulate internal symbolic logic representations, nor do they possess genuine comprehension in the human sense. Nevertheless, this simulation can be highly effective in practice. LLMs utilising CoT achieve strong performance on various reasoning and problem-solving benchmarks (e.g., arithmetic reasoning, commonsense reasoning); their simulated reasoning is often sufficient to solve complex tasks correctly. The key point for philosophical understanding, then, is that the outputs generated via CoT – these simulated arguments – are precisely the type of structured text that facilitates the mapping of dependence relations central to the account of understanding adopted here. These generated arguments can be complex and detailed, providing substantive material for philosophical analysis [Placeholder for examples of complex CoT philosophical arguments]. Moreover, the utility of these outputs parallels our engagement with human-generated philosophical texts; an argument need not be perfectly sound, nor entirely accepted by the reader, to enhance their understanding. Evaluating an argument, even one containing flawed steps, prompts the reader to critically examine the purported dependencies and refine their own representation of the conceptual network. This process of refinement seems equally applicable when the text being evaluated is a CoT output from an LLM. The effectiveness of this simulation is further supported by the observation that CoT capabilities emerge strongly with increased model scale (parameters numbering in the hundreds of billions or trillions) and are enhanced by refinement techniques such as Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which serve to improve the coherence, reliability, and instruction-following behaviour of the generated reasoning. Therefore, CoT connects directly to established philosophical methodology. It facilitates the generation of outputs structured like arguments, mirroring the sequential steps of premise-setting, inference, and conclusion characteristic of philosophical discourse. This structure, providing a potential pathway rather than just an endpoint, lends itself to analysis. It offers an artefact for philosophical scrutiny whose value derives from its effectiveness as a simulation, irrespective of its origin in a non-understanding source. The crucial outcome of these developments is that LLMs can produce highly effective simulations of reasoning. While they lack genuine understanding, their ability to generate step-by-step outputs significantly improves their performance on complex tasks. Models utilising CoT or similar structured reasoning approaches demonstrate strong results on various benchmarks requiring logical inference, arithmetic calculation, or commonsense reasoning. For many problem-solving purposes, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions. Understanding how this simulation works is key. As established, LLMs operate by predicting the next token based on preceding context, using the Transformer architecture's self-attention mechanism to weigh the relevance of earlier tokens. They do not possess internal logic modules or conceptual understanding. Their ability to generate coherent reasoning chains stems from learning statistical patterns in their vast training data, which contains countless examples of human arguments, explanations, and step-by-step procedures. When prompted appropriately, the model activates these learned patterns, generating sequences that mimic the form of reasoning. The CoT process leverages the model's architecture effectively. By prompting the model to generate intermediate steps, these steps are added to the context window. The self-attention mechanism can then 'refer back' to these explicit steps when generating subsequent tokens, maintaining coherence and allowing later steps to build upon earlier ones. This turns the context window into a kind of 'scratchpad', enabling the simulation of a structured thought process. The quality and reliability of this simulation are further enhanced by the sheer scale of modern models and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which train the model to better follow instructions and produce outputs aligned with human judgments of coherence and accuracy. [^1]: There are of course degrees of significance, some philosophical works have a profound impact on the philosophical understanding of large swathes of people (Gettier, Kant etc.). [^2]: mention butlin, buckner and m etc. [^3]: Mention also the way they are trained being very different to how human's learn. * Scrap?? Perhaps suprisingly, models fared much better at this sort of problem when they were told to ‘Think step by step’, [leading them to….] This guidance can come from prompting techniques like Chain-of-Thought (CoT). CoT prompting involves instructing the model – either by providing examples (few-shot) or by using direct commands like 'Let's think step by step' (zero-shot) – to articulate intermediate steps before providing a final answer. Instead of generating only a conclusion, the model is steered towards producing a sequence that outlines a potential reasoning process, such as "Step 1: Identify the letters... Step 2: Count the 'r's... Step 3: State the total." This technique works by leveraging the standard mechanisms of autoregressive LLMs. Recall that these models generate text token-by-token, with each prediction conditioned on the preceding context, and use self-attention to weigh the relevance of earlier parts of that context. When CoT prompting is used, the initial instruction biases the model to start generating tokens that represent the first step of a reasoning process. These generated tokens are then added to the ongoing context. For the next prediction, the model attends to this augmented context, including the first step, which influences it to generate tokens representing the second step, and so on. The intermediate steps thus become part of the conditioning context for subsequent steps, effectively creating an explicit 'scratchpad' within the model's context window that guides the generation towards a structured, step-by-step output. Their ability to produce coherent reasoning chains derives from learning complex statistical patterns in their training data, which includes vast amounts of text exhibiting step-by-step explanations, arguments, and problem-solving procedures. CoT prompting activates these learned patterns, causing the model to generate text that mimics, or simulates the form and sequence of reasoning it has observed, without actually understanding the logical or conceptual relationships involved. The quality and reliability of this simulation depend heavily on factors like model scale – CoT effects are much more pronounced in very large models – and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which improve the model's ability to follow instructions and generate outputs that align with human judgments of coherence and accuracy. Many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) seem to integrate these structured reasoning capabilities more deeply, often producing step-by-step outputs without explicit CoT prompts. Despite being a simulation, this capability is highly effective. Models guided by CoT principles demonstrate significantly improved performance on a wide range of benchmarks requiring multi-step reasoning, such as arithmetic word problems, commonsense inference, and symbolic manipulation tasks. For many problem-solving purposes across various domains, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions where simpler prompting methods fail. This effectiveness connects directly to the potential for enhancing philosophical understanding. The simulation of reasoning allows LLMs to generate outputs structured like arguments, presenting sequences of claims linked by inferential markers and intermediate steps. As established earlier, it is precisely this type of structured output, rather than isolated pronouncements, that typically facilitates the mapping of dependence relations central to philosophical understanding. These simulated arguments can be complex and detailed, offering substantive textual material for analysis. [Placeholder for examples of complex CoT philosophical arguments]. Furthermore, the philosophical utility of these simulated arguments mirrors our engagement with human-authored texts. A philosophical argument need not be perfectly sound, nor its conclusions fully endorsed, for it to be intellectually valuable. Engaging with an argument – identifying its premises, scrutinising its steps, evaluating its inferences, even if flawed – prompts the reader to critically examine the purported dependencies, clarify concepts, and potentially revise or refine their own representation of the relevant conceptual network. The same potential for refinement exists when engaging with the structured output of an LLM. The fact that the reasoning is simulated does not preclude the output from serving as a useful artefact for philosophical reflection, analysis, and the enhancement of the user's own understanding, particularly given the 'epistemically undemanding' nature of understanding adopted in §1. 4. Enhancing Understanding via Chain-of-Thought Outputs The structured outputs generated via CoT prompting can directly enhance philosophical understanding as defined by the Dellsén framework – the accurate and comprehensive representation of dependence networks. The value lies not in assuming the LLM 'understands' the dependencies it articulates, but in how the generated text can prompt refinements in the user's representation. • Improving Accuracy: By presenting a step-by-step derivation, CoT outputs allow the user to scrutinise the purported dependence relations involved. If the LLM generates an argument linking concept A to concept C via an intermediate step B, the user can evaluate the claimed dependence of C on B, and B on A. Identifying a weak or fallacious link in the chain prompts the user to correct their own mental model, potentially removing an inaccurately represented dependency. Conversely, a valid step might reinforce or confirm a dependency the user held tentatively. • Increasing Comprehensiveness: The intermediate steps generated by CoT can introduce new nodes or relations into the user's dependence network. The model might highlight a conceptual distinction, an implicit premise, or a relevant factor that the user had not previously considered, thereby expanding the scope of their representation. It might articulate dependencies between elements that the user had considered disparate. Furthermore, by tracing a line of reasoning, CoT can help delineate negative dependencies – showing, for instance, why a certain conclusion does not follow from a given premise under a specific interpretation, thus refining the boundaries of the represented network. • Clarifying Dependence Types: Philosophical understanding involves grasping how things hang together. The step-by-step structure can help disambiguate the nature of the proposed dependencies – are they logical, causal, conceptual, constitutive, or mereological? While the LLM might not explicitly label them, the context provided by the intermediate steps can allow the user to better infer the type of relationship being posited, leading to a more precise representation. Crucially, within the Dellsén framework, this enhancement does not require the user to have justification for, or even belief in, every proposition generated by the LLM. The framework is "epistemically undemanding" in this sense. The value derives from the potential of the generated structure to prompt a re-evaluation and refinement of the user's existing representation, leading to a potentially more accurate or comprehensive map of the relevant dependence network, even if parts of the generated chain are ultimately rejected upon reflection. Scrap?? This distinction between the system's internal state and its output's potential function resonates with the analysis by Butlin and Viebahn (forthcoming) regarding AI assertion. They argue that for an output to be an assertion, it must possess a 'descriptive function' – the function of conveying information to a consumer. Crucially, they contend that pre-trained LLMs, optimised solely to predict likely word sequences, produce outputs lacking this descriptive function. Their outputs mimic assertion form, but their function is merely sequence completion, not information conveyance. However, LLMs fine-tuned specifically for 'groundedness' or 'correctness', like LaMDA or Sparrow, undergo training that selects for accuracy and alignment with known sources. This fine-tuning process, Butlin and Viebahn suggest, means the outputs of such systems may acquire a descriptive function, serving the purpose of providing trustworthy information, even if the systems themselves fail other conditions for genuine assertion, such as sanctionability. This highlights that an output's functional properties, like potentially having a descriptive function, can be distinct from whether the generating system possesses understanding . One could argue that the origin of the text, whether from a thinking being or a non-understanding system, matters less than its content and its ability to make the human reader think philosophically. To take this argument further, one might use a concept like 'descriptive functions', possibly drawing on work by figures like Butler [Citation needed if specific work is intended]. The basic idea here is that even if an LLM output is not backed by the usual kind of authorial intention or understanding, the text itself can still function descriptively. It produces outputs designed to be useful, claiming to offer descriptions – perhaps unstated descriptions – of the world or parts of it. These outputs, because of their structure and content, could potentially contain or lead to philosophical insights, regardless of their source. This view, however, faces its own problems. A likely counter-argument might use the analogy of infinite monkeys eventually producing Shakespeare; if LLM outputs are just statistically generated sequences, why should their descriptive functions be taken seriously as potential sources of philosophical insight? This objection highlights the need for more study into why LLM-generated text, seen as having descriptive functions, might be considered important for knowledge or able to contribute meaningfully to philosophical discussion. Dealing with this requires a closer look at how LLMs work and the exact nature of these proposed descriptive functions. Historically, the capacity of LLMs to provide reasoned arguments has been subject to doubt. Earlier models, for instance, often produced outputs where the reasoning presented was weak, inconsistent, or seemingly confabulated [Optional: Add brief examples here]. The development of prompting techniques such as Chain-of-Thought (CoT), however, marked a significant shift in approach. CoT was designed specifically to elicit intermediate reasoning steps from LLMs, thereby improving performance on tasks requiring sequential reasoning processes. Common methods for eliciting CoT include few-shot prompting – providing examples of step-by-step reasoning within the prompt – and zero-shot prompting, which uses triggers like the phrase 'Let's think step by step'. Furthermore, many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) appear to incorporate CoT-like structured reasoning capabilities more implicitly; they often generate step-by-step explanations or arguments without explicit CoT triggers, suggesting this methodology is now somewhat automated or integrated via system prompts or subsequent fine-tuning stages. Technically, CoT leverages the underlying architecture of autoregressive Transformers. As detailed previously, these models predict the next token based upon the preceding sequence, utilising self-attention mechanisms to weigh the importance of different elements within the context. CoT prompting guides this generative process: instead of predicting the final answer directly, the model is steered towards generating tokens which represent intermediate steps in a reasoning sequence. Each generated step is then appended to the context. Consequently, the model's attention mechanism considers this augmented context – including the steps generated thus far – when predicting the subsequent token. This iterative process allows the model to condition later parts of its output on earlier parts, effectively employing the generated text as an explicit 'scratchpad' to manage the simulated reasoning process. It must be acknowledged that this process constitutes a simulation of reasoning, driven by the model learning statistical patterns from vast quantities of text data that include examples of arguments and explanations. LLMs do not manipulate internal symbolic logic representations, nor do they possess genuine comprehension in the human sense. Nevertheless, this simulation can be highly effective in practice. LLMs utilising CoT achieve strong performance on various reasoning and problem-solving benchmarks (e.g., arithmetic reasoning, commonsense reasoning); their simulated reasoning is often sufficient to solve complex tasks correctly. The key point for philosophical understanding, then, is that the outputs generated via CoT – these simulated arguments – are precisely the type of structured text that facilitates the mapping of dependence relations central to the account of understanding adopted here. These generated arguments can be complex and detailed, providing substantive material for philosophical analysis [Placeholder for examples of complex CoT philosophical arguments]. Moreover, the utility of these outputs parallels our engagement with human-generated philosophical texts; an argument need not be perfectly sound, nor entirely accepted by the reader, to enhance their understanding. Evaluating an argument, even one containing flawed steps, prompts the reader to critically examine the purported dependencies and refine their own representation of the conceptual network. This process of refinement seems equally applicable when the text being evaluated is a CoT output from an LLM. The effectiveness of this simulation is further supported by the observation that CoT capabilities emerge strongly with increased model scale (parameters numbering in the hundreds of billions or trillions) and are enhanced by refinement techniques such as Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which serve to improve the coherence, reliability, and instruction-following behaviour of the generated reasoning. Therefore, CoT connects directly to established philosophical methodology. It facilitates the generation of outputs structured like arguments, mirroring the sequential steps of premise-setting, inference, and conclusion characteristic of philosophical discourse. This structure, providing a potential pathway rather than just an endpoint, lends itself to analysis. It offers an artefact for philosophical scrutiny whose value derives from its effectiveness as a simulation, irrespective of its origin in a non-understanding source. The crucial outcome of these developments is that LLMs can produce highly effective simulations of reasoning. While they lack genuine understanding, their ability to generate step-by-step outputs significantly improves their performance on complex tasks. Models utilising CoT or similar structured reasoning approaches demonstrate strong results on various benchmarks requiring logical inference, arithmetic calculation, or commonsense reasoning. For many problem-solving purposes, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions. Understanding how this simulation works is key. As established, LLMs operate by predicting the next token based on preceding context, using the Transformer architecture's self-attention mechanism to weigh the relevance of earlier tokens. They do not possess internal logic modules or conceptual understanding. Their ability to generate coherent reasoning chains stems from learning statistical patterns in their vast training data, which contains countless examples of human arguments, explanations, and step-by-step procedures. When prompted appropriately, the model activates these learned patterns, generating sequences that mimic the form of reasoning. The CoT process leverages the model's architecture effectively. By prompting the model to generate intermediate steps, these steps are added to the context window. The self-attention mechanism can then 'refer back' to these explicit steps when generating subsequent tokens, maintaining coherence and allowing later steps to build upon earlier ones. This turns the context window into a kind of 'scratchpad', enabling the simulation of a structured thought process. The quality and reliability of this simulation are further enhanced by the sheer scale of modern models and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which train the model to better follow instructions and produce outputs aligned with human judgments of coherence and accuracy. [^1]: There are of course degrees of significance, some philosophical works have a profound impact on the philosophical understanding of large swathes of people (Gettier, Kant etc.). [^2]: mention butlin, buckner and m etc. [^3]: Mention also the way they are trained being very different to how human's learn. # Scrap?? Perhaps suprisingly, models fared much better at this sort of problem when they were told to ‘Think step by step’, [leading them to….] This guidance can come from prompting techniques like Chain-of-Thought (CoT). CoT prompting involves instructing the model – either by providing examples (few-shot) or by using direct commands like 'Let's think step by step' (zero-shot) – to articulate intermediate steps before providing a final answer. Instead of generating only a conclusion, the model is steered towards producing a sequence that outlines a potential reasoning process, such as "Step 1: Identify the letters... Step 2: Count the 'r's... Step 3: State the total." This technique works by leveraging the standard mechanisms of autoregressive LLMs. Recall that these models generate text token-by-token, with each prediction conditioned on the preceding context, and use self-attention to weigh the relevance of earlier parts of that context. When CoT prompting is used, the initial instruction biases the model to start generating tokens that represent the first step of a reasoning process. These generated tokens are then added to the ongoing context. For the next prediction, the model attends to this augmented context, including the first step, which influences it to generate tokens representing the second step, and so on. The intermediate steps thus become part of the conditioning context for subsequent steps, effectively creating an explicit 'scratchpad' within the model's context window that guides the generation towards a structured, step-by-step output. Their ability to produce coherent reasoning chains derives from learning complex statistical patterns in their training data, which includes vast amounts of text exhibiting step-by-step explanations, arguments, and problem-solving procedures. CoT prompting activates these learned patterns, causing the model to generate text that mimics, or simulates the form and sequence of reasoning it has observed, without actually understanding the logical or conceptual relationships involved. The quality and reliability of this simulation depend heavily on factors like model scale – CoT effects are much more pronounced in very large models – and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which improve the model's ability to follow instructions and generate outputs that align with human judgments of coherence and accuracy. Many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) seem to integrate these structured reasoning capabilities more deeply, often producing step-by-step outputs without explicit CoT prompts. Despite being a simulation, this capability is highly effective. Models guided by CoT principles demonstrate significantly improved performance on a wide range of benchmarks requiring multi-step reasoning, such as arithmetic word problems, commonsense inference, and symbolic manipulation tasks. For many problem-solving purposes across various domains, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions where simpler prompting methods fail. This effectiveness connects directly to the potential for enhancing philosophical understanding. The simulation of reasoning allows LLMs to generate outputs structured like arguments, presenting sequences of claims linked by inferential markers and intermediate steps. As established earlier, it is precisely this type of structured output, rather than isolated pronouncements, that typically facilitates the mapping of dependence relations central to philosophical understanding. These simulated arguments can be complex and detailed, offering substantive textual material for analysis. [Placeholder for examples of complex CoT philosophical arguments]. Furthermore, the philosophical utility of these simulated arguments mirrors our engagement with human-authored texts. A philosophical argument need not be perfectly sound, nor its conclusions fully endorsed, for it to be intellectually valuable. Engaging with an argument – identifying its premises, scrutinising its steps, evaluating its inferences, even if flawed – prompts the reader to critically examine the purported dependencies, clarify concepts, and potentially revise or refine their own representation of the relevant conceptual network. The same potential for refinement exists when engaging with the structured output of an LLM. The fact that the reasoning is simulated does not preclude the output from serving as a useful artefact for philosophical reflection, analysis, and the enhancement of the user's own understanding, particularly given the 'epistemically undemanding' nature of understanding adopted in §1. 4. Enhancing Understanding via Chain-of-Thought Outputs The structured outputs generated via CoT prompting can directly enhance philosophical understanding as defined by the Dellsén framework – the accurate and comprehensive representation of dependence networks. The value lies not in assuming the LLM 'understands' the dependencies it articulates, but in how the generated text can prompt refinements in the user's representation. • Improving Accuracy: By presenting a step-by-step derivation, CoT outputs allow the user to scrutinise the purported dependence relations involved. If the LLM generates an argument linking concept A to concept C via an intermediate step B, the user can evaluate the claimed dependence of C on B, and B on A. Identifying a weak or fallacious link in the chain prompts the user to correct their own mental model, potentially removing an inaccurately represented dependency. Conversely, a valid step might reinforce or confirm a dependency the user held tentatively. • Increasing Comprehensiveness: The intermediate steps generated by CoT can introduce new nodes or relations into the user's dependence network. The model might highlight a conceptual distinction, an implicit premise, or a relevant factor that the user had not previously considered, thereby expanding the scope of their representation. It might articulate dependencies between elements that the user had considered disparate. Furthermore, by tracing a line of reasoning, CoT can help delineate negative dependencies – showing, for instance, why a certain conclusion does not follow from a given premise under a specific interpretation, thus refining the boundaries of the represented network. • Clarifying Dependence Types: Philosophical understanding involves grasping how things hang together. The step-by-step structure can help disambiguate the nature of the proposed dependencies – are they logical, causal, conceptual, constitutive, or mereological? While the LLM might not explicitly label them, the context provided by the intermediate steps can allow the user to better infer the type of relationship being posited, leading to a more precise representation. Crucially, within the Dellsén framework, this enhancement does not require the user to have justification for, or even belief in, every proposition generated by the LLM. The framework is "epistemically undemanding" in this sense. The value derives from the potential of the generated structure to prompt a re-evaluation and refinement of the user's existing representation, leading to a potentially more accurate or comprehensive map of the relevant dependence network, even if parts of the generated chain are ultimately rejected upon reflection. ## Scrap?? - This distinction between the system's internal state and its output's potential function resonates with the analysis by Butlin and Viebahn (forthcoming) regarding AI assertion. - They argue that for an output to be an assertion, it must possess a 'descriptive function' – the function of conveying information to a consumer. - Crucially, they contend that pre-trained LLMs, optimised solely to predict likely word sequences, produce outputs lacking this descriptive function. - Their outputs mimic assertion form, but their function is merely sequence completion, not information conveyance. - However, LLMs fine-tuned specifically for 'groundedness' or 'correctness', like LaMDA or Sparrow, undergo training that selects for accuracy and alignment with known sources. - This fine-tuning process, Butlin and Viebahn suggest, means the outputs of such systems may acquire a descriptive function, serving the purpose of providing trustworthy information, even if the systems themselves fail other conditions for genuine assertion, such as sanctionability. - This highlights that an output's functional properties, like potentially having a descriptive function, can be distinct from whether the generating system possesses understanding . One could argue that the origin of the text, whether from a thinking being or a non-understanding system, matters less than its content and its ability to make the human reader think philosophically. To take this argument further, one might use a concept like 'descriptive functions', possibly drawing on work by figures like Butler [Citation needed if specific work is intended]. The basic idea here is that even if an LLM output is not backed by the usual kind of authorial intention or understanding, the text itself can still function descriptively. It produces outputs designed to be useful, claiming to offer descriptions – perhaps unstated descriptions – of the world or parts of it. These outputs, because of their structure and content, could potentially contain or lead to philosophical insights, regardless of their source. This view, however, faces its own problems. A likely counter-argument might use the analogy of infinite monkeys eventually producing Shakespeare; if LLM outputs are just statistically generated sequences, why should their descriptive functions be taken seriously as potential sources of philosophical insight? This objection highlights the need for more study into why LLM-generated text, seen as having descriptive functions, might be considered important for knowledge or able to contribute meaningfully to philosophical discussion. Dealing with this requires a closer look at how LLMs work and the exact nature of these proposed descriptive functions. Historically, the capacity of LLMs to provide reasoned arguments has been subject to doubt. Earlier models, for instance, often produced outputs where the reasoning presented was weak, inconsistent, or seemingly confabulated [Optional: Add brief examples here]. The development of prompting techniques such as Chain-of-Thought (CoT), however, marked a significant shift in approach. CoT was designed specifically to elicit intermediate reasoning steps from LLMs, thereby improving performance on tasks requiring sequential reasoning processes. Common methods for eliciting CoT include few-shot prompting – providing examples of step-by-step reasoning within the prompt – and zero-shot prompting, which uses triggers like the phrase 'Let's think step by step'. Furthermore, many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) appear to incorporate CoT-like structured reasoning capabilities more implicitly; they often generate step-by-step explanations or arguments without explicit CoT triggers, suggesting this methodology is now somewhat automated or integrated via system prompts or subsequent fine-tuning stages. Technically, CoT leverages the underlying architecture of autoregressive Transformers. As detailed previously, these models predict the next token based upon the preceding sequence, utilising self-attention mechanisms to weigh the importance of different elements within the context. CoT prompting guides this generative process: instead of predicting the final answer directly, the model is steered towards generating tokens which represent intermediate steps in a reasoning sequence. Each generated step is then appended to the context. Consequently, the model's attention mechanism considers this augmented context – including the steps generated thus far – when predicting the subsequent token. This iterative process allows the model to condition later parts of its output on earlier parts, effectively employing the generated text as an explicit 'scratchpad' to manage the simulated reasoning process. It must be acknowledged that this process constitutes a simulation of reasoning, driven by the model learning statistical patterns from vast quantities of text data that include examples of arguments and explanations. LLMs do not manipulate internal symbolic logic representations, nor do they possess genuine comprehension in the human sense. Nevertheless, this simulation can be highly effective in practice. LLMs utilising CoT achieve strong performance on various reasoning and problem-solving benchmarks (e.g., arithmetic reasoning, commonsense reasoning); their simulated reasoning is often sufficient to solve complex tasks correctly. The key point for philosophical understanding, then, is that the outputs generated via CoT – these simulated arguments – are precisely the type of structured text that facilitates the mapping of dependence relations central to the account of understanding adopted here. These generated arguments can be complex and detailed, providing substantive material for philosophical analysis [Placeholder for examples of complex CoT philosophical arguments]. Moreover, the utility of these outputs parallels our engagement with human-generated philosophical texts; an argument need not be perfectly sound, nor entirely accepted by the reader, to enhance their understanding. Evaluating an argument, even one containing flawed steps, prompts the reader to critically examine the purported dependencies and refine their own representation of the conceptual network. This process of refinement seems equally applicable when the text being evaluated is a CoT output from an LLM. The effectiveness of this simulation is further supported by the observation that CoT capabilities emerge strongly with increased model scale (parameters numbering in the hundreds of billions or trillions) and are enhanced by refinement techniques such as Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which serve to improve the coherence, reliability, and instruction-following behaviour of the generated reasoning. Therefore, CoT connects directly to established philosophical methodology. It facilitates the generation of outputs structured like arguments, mirroring the sequential steps of premise-setting, inference, and conclusion characteristic of philosophical discourse. This structure, providing a potential pathway rather than just an endpoint, lends itself to analysis. It offers an artefact for philosophical scrutiny whose value derives from its effectiveness as a simulation, irrespective of its origin in a non-understanding source. The crucial outcome of these developments is that LLMs can produce highly effective simulations of reasoning. While they lack genuine understanding, their ability to generate step-by-step outputs significantly improves their performance on complex tasks. Models utilising CoT or similar structured reasoning approaches demonstrate strong results on various benchmarks requiring logical inference, arithmetic calculation, or commonsense reasoning. For many problem-solving purposes, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions. Understanding how this simulation works is key. As established, LLMs operate by predicting the next token based on preceding context, using the Transformer architecture's self-attention mechanism to weigh the relevance of earlier tokens. They do not possess internal logic modules or conceptual understanding. Their ability to generate coherent reasoning chains stems from learning statistical patterns in their vast training data, which contains countless examples of human arguments, explanations, and step-by-step procedures. When prompted appropriately, the model activates these learned patterns, generating sequences that mimic the form of reasoning. The CoT process leverages the model's architecture effectively. By prompting the model to generate intermediate steps, these steps are added to the context window. The self-attention mechanism can then 'refer back' to these explicit steps when generating subsequent tokens, maintaining coherence and allowing later steps to build upon earlier ones. This turns the context window into a kind of 'scratchpad', enabling the simulation of a structured thought process. The quality and reliability of this simulation are further enhanced by the sheer scale of modern models and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which train the model to better follow instructions and produce outputs aligned with human judgments of coherence and accuracy. [^1]: There are of course degrees of significance, some philosophical works have a profound impact on the philosophical understanding of large swathes of people (Gettier, Kant etc.). [^2]: mention butlin, buckner and m etc. [^3]: Mention also the way they are trained being very different to how human's learn. ** # Scrap?? Perhaps suprisingly, models fared much better at this sort of problem when they were told to ‘Think step by step’, [leading them to….] This guidance can come from prompting techniques like Chain-of-Thought (CoT). CoT prompting involves instructing the model – either by providing examples (few-shot) or by using direct commands like 'Let's think step by step' (zero-shot) – to articulate intermediate steps before providing a final answer. Instead of generating only a conclusion, the model is steered towards producing a sequence that outlines a potential reasoning process, such as "Step 1: Identify the letters... Step 2: Count the 'r's... Step 3: State the total." This technique works by leveraging the standard mechanisms of autoregressive LLMs. Recall that these models generate text token-by-token, with each prediction conditioned on the preceding context, and use self-attention to weigh the relevance of earlier parts of that context. When CoT prompting is used, the initial instruction biases the model to start generating tokens that represent the first step of a reasoning process. These generated tokens are then added to the ongoing context. For the next prediction, the model attends to this augmented context, including the first step, which influences it to generate tokens representing the second step, and so on. The intermediate steps thus become part of the conditioning context for subsequent steps, effectively creating an explicit 'scratchpad' within the model's context window that guides the generation towards a structured, step-by-step output. Their ability to produce coherent reasoning chains derives from learning complex statistical patterns in their training data, which includes vast amounts of text exhibiting step-by-step explanations, arguments, and problem-solving procedures. CoT prompting activates these learned patterns, causing the model to generate text that mimics, or simulates the form and sequence of reasoning it has observed, without actually understanding the logical or conceptual relationships involved. The quality and reliability of this simulation depend heavily on factors like model scale – CoT effects are much more pronounced in very large models – and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which improve the model's ability to follow instructions and generate outputs that align with human judgments of coherence and accuracy. Many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) seem to integrate these structured reasoning capabilities more deeply, often producing step-by-step outputs without explicit CoT prompts. Despite being a simulation, this capability is highly effective. Models guided by CoT principles demonstrate significantly improved performance on a wide range of benchmarks requiring multi-step reasoning, such as arithmetic word problems, commonsense inference, and symbolic manipulation tasks. For many problem-solving purposes across various domains, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions where simpler prompting methods fail. This effectiveness connects directly to the potential for enhancing philosophical understanding. The simulation of reasoning allows LLMs to generate outputs structured like arguments, presenting sequences of claims linked by inferential markers and intermediate steps. As established earlier, it is precisely this type of structured output, rather than isolated pronouncements, that typically facilitates the mapping of dependence relations central to philosophical understanding. These simulated arguments can be complex and detailed, offering substantive textual material for analysis. [Placeholder for examples of complex CoT philosophical arguments]. Furthermore, the philosophical utility of these simulated arguments mirrors our engagement with human-authored texts. A philosophical argument need not be perfectly sound, nor its conclusions fully endorsed, for it to be intellectually valuable. Engaging with an argument – identifying its premises, scrutinising its steps, evaluating its inferences, even if flawed – prompts the reader to critically examine the purported dependencies, clarify concepts, and potentially revise or refine their own representation of the relevant conceptual network. The same potential for refinement exists when engaging with the structured output of an LLM. The fact that the reasoning is simulated does not preclude the output from serving as a useful artefact for philosophical reflection, analysis, and the enhancement of the user's own understanding, particularly given the 'epistemically undemanding' nature of understanding adopted in §1. 4. Enhancing Understanding via Chain-of-Thought Outputs The structured outputs generated via CoT prompting can directly enhance philosophical understanding as defined by the Dellsén framework – the accurate and comprehensive representation of dependence networks. The value lies not in assuming the LLM 'understands' the dependencies it articulates, but in how the generated text can prompt refinements in the user's representation. • Improving Accuracy: By presenting a step-by-step derivation, CoT outputs allow the user to scrutinise the purported dependence relations involved. If the LLM generates an argument linking concept A to concept C via an intermediate step B, the user can evaluate the claimed dependence of C on B, and B on A. Identifying a weak or fallacious link in the chain prompts the user to correct their own mental model, potentially removing an inaccurately represented dependency. Conversely, a valid step might reinforce or confirm a dependency the user held tentatively. • Increasing Comprehensiveness: The intermediate steps generated by CoT can introduce new nodes or relations into the user's dependence network. The model might highlight a conceptual distinction, an implicit premise, or a relevant factor that the user had not previously considered, thereby expanding the scope of their representation. It might articulate dependencies between elements that the user had considered disparate. Furthermore, by tracing a line of reasoning, CoT can help delineate negative dependencies – showing, for instance, why a certain conclusion does not follow from a given premise under a specific interpretation, thus refining the boundaries of the represented network. • Clarifying Dependence Types: Philosophical understanding involves grasping how things hang together. The step-by-step structure can help disambiguate the nature of the proposed dependencies – are they logical, causal, conceptual, constitutive, or mereological? While the LLM might not explicitly label them, the context provided by the intermediate steps can allow the user to better infer the type of relationship being posited, leading to a more precise representation. Crucially, within the Dellsén framework, this enhancement does not require the user to have justification for, or even belief in, every proposition generated by the LLM. The framework is "epistemically undemanding" in this sense. The value derives from the potential of the generated structure to prompt a re-evaluation and refinement of the user's existing representation, leading to a potentially more accurate or comprehensive map of the relevant dependence network, even if parts of the generated chain are ultimately rejected upon reflection. ## Scrap?? - This distinction between the system's internal state and its output's potential function resonates with the analysis by Butlin and Viebahn (forthcoming) regarding AI assertion. - They argue that for an output to be an assertion, it must possess a 'descriptive function' – the function of conveying information to a consumer. - Crucially, they contend that pre-trained LLMs, optimised solely to predict likely word sequences, produce outputs lacking this descriptive function. - Their outputs mimic assertion form, but their function is merely sequence completion, not information conveyance. - However, LLMs fine-tuned specifically for 'groundedness' or 'correctness', like LaMDA or Sparrow, undergo training that selects for accuracy and alignment with known sources. - This fine-tuning process, Butlin and Viebahn suggest, means the outputs of such systems may acquire a descriptive function, serving the purpose of providing trustworthy information, even if the systems themselves fail other conditions for genuine assertion, such as sanctionability. - This highlights that an output's functional properties, like potentially having a descriptive function, can be distinct from whether the generating system possesses understanding . One could argue that the origin of the text, whether from a thinking being or a non-understanding system, matters less than its content and its ability to make the human reader think philosophically. To take this argument further, one might use a concept like 'descriptive functions', possibly drawing on work by figures like Butler [Citation needed if specific work is intended]. The basic idea here is that even if an LLM output is not backed by the usual kind of authorial intention or understanding, the text itself can still function descriptively. It produces outputs designed to be useful, claiming to offer descriptions – perhaps unstated descriptions – of the world or parts of it. These outputs, because of their structure and content, could potentially contain or lead to philosophical insights, regardless of their source. This view, however, faces its own problems. A likely counter-argument might use the analogy of infinite monkeys eventually producing Shakespeare; if LLM outputs are just statistically generated sequences, why should their descriptive functions be taken seriously as potential sources of philosophical insight? This objection highlights the need for more study into why LLM-generated text, seen as having descriptive functions, might be considered important for knowledge or able to contribute meaningfully to philosophical discussion. Dealing with this requires a closer look at how LLMs work and the exact nature of these proposed descriptive functions. Historically, the capacity of LLMs to provide reasoned arguments has been subject to doubt. Earlier models, for instance, often produced outputs where the reasoning presented was weak, inconsistent, or seemingly confabulated [Optional: Add brief examples here]. The development of prompting techniques such as Chain-of-Thought (CoT), however, marked a significant shift in approach. CoT was designed specifically to elicit intermediate reasoning steps from LLMs, thereby improving performance on tasks requiring sequential reasoning processes. Common methods for eliciting CoT include few-shot prompting – providing examples of step-by-step reasoning within the prompt – and zero-shot prompting, which uses triggers like the phrase 'Let's think step by step'. Furthermore, many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) appear to incorporate CoT-like structured reasoning capabilities more implicitly; they often generate step-by-step explanations or arguments without explicit CoT triggers, suggesting this methodology is now somewhat automated or integrated via system prompts or subsequent fine-tuning stages. Technically, CoT leverages the underlying architecture of autoregressive Transformers. As detailed previously, these models predict the next token based upon the preceding sequence, utilising self-attention mechanisms to weigh the importance of different elements within the context. CoT prompting guides this generative process: instead of predicting the final answer directly, the model is steered towards generating tokens which represent intermediate steps in a reasoning sequence. Each generated step is then appended to the context. Consequently, the model's attention mechanism considers this augmented context – including the steps generated thus far – when predicting the subsequent token. This iterative process allows the model to condition later parts of its output on earlier parts, effectively employing the generated text as an explicit 'scratchpad' to manage the simulated reasoning process. It must be acknowledged that this process constitutes a simulation of reasoning, driven by the model learning statistical patterns from vast quantities of text data that include examples of arguments and explanations. LLMs do not manipulate internal symbolic logic representations, nor do they possess genuine comprehension in the human sense. Nevertheless, this simulation can be highly effective in practice. LLMs utilising CoT achieve strong performance on various reasoning and problem-solving benchmarks (e.g., arithmetic reasoning, commonsense reasoning); their simulated reasoning is often sufficient to solve complex tasks correctly. The key point for philosophical understanding, then, is that the outputs generated via CoT – these simulated arguments – are precisely the type of structured text that facilitates the mapping of dependence relations central to the account of understanding adopted here. These generated arguments can be complex and detailed, providing substantive material for philosophical analysis [Placeholder for examples of complex CoT philosophical arguments]. Moreover, the utility of these outputs parallels our engagement with human-generated philosophical texts; an argument need not be perfectly sound, nor entirely accepted by the reader, to enhance their understanding. Evaluating an argument, even one containing flawed steps, prompts the reader to critically examine the purported dependencies and refine their own representation of the conceptual network. This process of refinement seems equally applicable when the text being evaluated is a CoT output from an LLM. The effectiveness of this simulation is further supported by the observation that CoT capabilities emerge strongly with increased model scale (parameters numbering in the hundreds of billions or trillions) and are enhanced by refinement techniques such as Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which serve to improve the coherence, reliability, and instruction-following behaviour of the generated reasoning. Therefore, CoT connects directly to established philosophical methodology. It facilitates the generation of outputs structured like arguments, mirroring the sequential steps of premise-setting, inference, and conclusion characteristic of philosophical discourse. This structure, providing a potential pathway rather than just an endpoint, lends itself to analysis. It offers an artefact for philosophical scrutiny whose value derives from its effectiveness as a simulation, irrespective of its origin in a non-understanding source. The crucial outcome of these developments is that LLMs can produce highly effective simulations of reasoning. While they lack genuine understanding, their ability to generate step-by-step outputs significantly improves their performance on complex tasks. Models utilising CoT or similar structured reasoning approaches demonstrate strong results on various benchmarks requiring logical inference, arithmetic calculation, or commonsense reasoning. For many problem-solving purposes, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions. Understanding how this simulation works is key. As established, LLMs operate by predicting the next token based on preceding context, using the Transformer architecture's self-attention mechanism to weigh the relevance of earlier tokens. They do not possess internal logic modules or conceptual understanding. Their ability to generate coherent reasoning chains stems from learning statistical patterns in their vast training data, which contains countless examples of human arguments, explanations, and step-by-step procedures. When prompted appropriately, the model activates these learned patterns, generating sequences that mimic the form of reasoning. The CoT process leverages the model's architecture effectively. By prompting the model to generate intermediate steps, these steps are added to the context window. The self-attention mechanism can then 'refer back' to these explicit steps when generating subsequent tokens, maintaining coherence and allowing later steps to build upon earlier ones. This turns the context window into a kind of 'scratchpad', enabling the simulation of a structured thought process. The quality and reliability of this simulation are further enhanced by the sheer scale of modern models and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which train the model to better follow instructions and produce outputs aligned with human judgments of coherence and accuracy. [^1]: There are of course degrees of significance, some philosophical works have a profound impact on the philosophical understanding of large swathes of people (Gettier, Kant etc.). [^2]: mention butlin, buckner and m etc. [^3]: Mention also the way they are trained being very different to how human's learn. # Scrap?? Perhaps suprisingly, models fared much better at this sort of problem when they were told to ‘Think step by step’, [leading them to….] This guidance can come from prompting techniques like Chain-of-Thought (CoT). CoT prompting involves instructing the model – either by providing examples (few-shot) or by using direct commands like 'Let's think step by step' (zero-shot) – to articulate intermediate steps before providing a final answer. Instead of generating only a conclusion, the model is steered towards producing a sequence that outlines a potential reasoning process, such as "Step 1: Identify the letters... Step 2: Count the 'r's... Step 3: State the total." This technique works by leveraging the standard mechanisms of autoregressive LLMs. Recall that these models generate text token-by-token, with each prediction conditioned on the preceding context, and use self-attention to weigh the relevance of earlier parts of that context. When CoT prompting is used, the initial instruction biases the model to start generating tokens that represent the first step of a reasoning process. These generated tokens are then added to the ongoing context. For the next prediction, the model attends to this augmented context, including the first step, which influences it to generate tokens representing the second step, and so on. The intermediate steps thus become part of the conditioning context for subsequent steps, effectively creating an explicit 'scratchpad' within the model's context window that guides the generation towards a structured, step-by-step output. Their ability to produce coherent reasoning chains derives from learning complex statistical patterns in their training data, which includes vast amounts of text exhibiting step-by-step explanations, arguments, and problem-solving procedures. CoT prompting activates these learned patterns, causing the model to generate text that mimics, or simulates the form and sequence of reasoning it has observed, without actually understanding the logical or conceptual relationships involved. The quality and reliability of this simulation depend heavily on factors like model scale – CoT effects are much more pronounced in very large models – and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which improve the model's ability to follow instructions and generate outputs that align with human judgments of coherence and accuracy. Many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) seem to integrate these structured reasoning capabilities more deeply, often producing step-by-step outputs without explicit CoT prompts. Despite being a simulation, this capability is highly effective. Models guided by CoT principles demonstrate significantly improved performance on a wide range of benchmarks requiring multi-step reasoning, such as arithmetic word problems, commonsense inference, and symbolic manipulation tasks. For many problem-solving purposes across various domains, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions where simpler prompting methods fail. This effectiveness connects directly to the potential for enhancing philosophical understanding. The simulation of reasoning allows LLMs to generate outputs structured like arguments, presenting sequences of claims linked by inferential markers and intermediate steps. As established earlier, it is precisely this type of structured output, rather than isolated pronouncements, that typically facilitates the mapping of dependence relations central to philosophical understanding. These simulated arguments can be complex and detailed, offering substantive textual material for analysis. [Placeholder for examples of complex CoT philosophical arguments]. Furthermore, the philosophical utility of these simulated arguments mirrors our engagement with human-authored texts. A philosophical argument need not be perfectly sound, nor its conclusions fully endorsed, for it to be intellectually valuable. Engaging with an argument – identifying its premises, scrutinising its steps, evaluating its inferences, even if flawed – prompts the reader to critically examine the purported dependencies, clarify concepts, and potentially revise or refine their own representation of the relevant conceptual network. The same potential for refinement exists when engaging with the structured output of an LLM. The fact that the reasoning is simulated does not preclude the output from serving as a useful artefact for philosophical reflection, analysis, and the enhancement of the user's own understanding, particularly given the 'epistemically undemanding' nature of understanding adopted in §1. 4. Enhancing Understanding via Chain-of-Thought Outputs The structured outputs generated via CoT prompting can directly enhance philosophical understanding as defined by the Dellsén framework – the accurate and comprehensive representation of dependence networks. The value lies not in assuming the LLM 'understands' the dependencies it articulates, but in how the generated text can prompt refinements in the user's representation. • Improving Accuracy: By presenting a step-by-step derivation, CoT outputs allow the user to scrutinise the purported dependence relations involved. If the LLM generates an argument linking concept A to concept C via an intermediate step B, the user can evaluate the claimed dependence of C on B, and B on A. Identifying a weak or fallacious link in the chain prompts the user to correct their own mental model, potentially removing an inaccurately represented dependency. Conversely, a valid step might reinforce or confirm a dependency the user held tentatively. • Increasing Comprehensiveness: The intermediate steps generated by CoT can introduce new nodes or relations into the user's dependence network. The model might highlight a conceptual distinction, an implicit premise, or a relevant factor that the user had not previously considered, thereby expanding the scope of their representation. It might articulate dependencies between elements that the user had considered disparate. Furthermore, by tracing a line of reasoning, CoT can help delineate negative dependencies – showing, for instance, why a certain conclusion does not follow from a given premise under a specific interpretation, thus refining the boundaries of the represented network. • Clarifying Dependence Types: Philosophical understanding involves grasping how things hang together. The step-by-step structure can help disambiguate the nature of the proposed dependencies – are they logical, causal, conceptual, constitutive, or mereological? While the LLM might not explicitly label them, the context provided by the intermediate steps can allow the user to better infer the type of relationship being posited, leading to a more precise representation. Crucially, within the Dellsén framework, this enhancement does not require the user to have justification for, or even belief in, every proposition generated by the LLM. The framework is "epistemically undemanding" in this sense. The value derives from the potential of the generated structure to prompt a re-evaluation and refinement of the user's existing representation, leading to a potentially more accurate or comprehensive map of the relevant dependence network, even if parts of the generated chain are ultimately rejected upon reflection. ## Scrap?? - This distinction between the system's internal state and its output's potential function resonates with the analysis by Butlin and Viebahn (forthcoming) regarding AI assertion. - They argue that for an output to be an assertion, it must possess a 'descriptive function' – the function of conveying information to a consumer. - Crucially, they contend that pre-trained LLMs, optimised solely to predict likely word sequences, produce outputs lacking this descriptive function. - Their outputs mimic assertion form, but their function is merely sequence completion, not information conveyance. - However, LLMs fine-tuned specifically for 'groundedness' or 'correctness', like LaMDA or Sparrow, undergo training that selects for accuracy and alignment with known sources. - This fine-tuning process, Butlin and Viebahn suggest, means the outputs of such systems may acquire a descriptive function, serving the purpose of providing trustworthy information, even if the systems themselves fail other conditions for genuine assertion, such as sanctionability. - This highlights that an output's functional properties, like potentially having a descriptive function, can be distinct from whether the generating system possesses understanding . One could argue that the origin of the text, whether from a thinking being or a non-understanding system, matters less than its content and its ability to make the human reader think philosophically. To take this argument further, one might use a concept like 'descriptive functions', possibly drawing on work by figures like Butler [Citation needed if specific work is intended]. The basic idea here is that even if an LLM output is not backed by the usual kind of authorial intention or understanding, the text itself can still function descriptively. It produces outputs designed to be useful, claiming to offer descriptions – perhaps unstated descriptions – of the world or parts of it. These outputs, because of their structure and content, could potentially contain or lead to philosophical insights, regardless of their source. This view, however, faces its own problems. A likely counter-argument might use the analogy of infinite monkeys eventually producing Shakespeare; if LLM outputs are just statistically generated sequences, why should their descriptive functions be taken seriously as potential sources of philosophical insight? This objection highlights the need for more study into why LLM-generated text, seen as having descriptive functions, might be considered important for knowledge or able to contribute meaningfully to philosophical discussion. Dealing with this requires a closer look at how LLMs work and the exact nature of these proposed descriptive functions. Historically, the capacity of LLMs to provide reasoned arguments has been subject to doubt. Earlier models, for instance, often produced outputs where the reasoning presented was weak, inconsistent, or seemingly confabulated [Optional: Add brief examples here]. The development of prompting techniques such as Chain-of-Thought (CoT), however, marked a significant shift in approach. CoT was designed specifically to elicit intermediate reasoning steps from LLMs, thereby improving performance on tasks requiring sequential reasoning processes. Common methods for eliciting CoT include few-shot prompting – providing examples of step-by-step reasoning within the prompt – and zero-shot prompting, which uses triggers like the phrase 'Let's think step by step'. Furthermore, many current advanced models (e.g., Gemini 2.5 Pro, ChatGPT o3) appear to incorporate CoT-like structured reasoning capabilities more implicitly; they often generate step-by-step explanations or arguments without explicit CoT triggers, suggesting this methodology is now somewhat automated or integrated via system prompts or subsequent fine-tuning stages. Technically, CoT leverages the underlying architecture of autoregressive Transformers. As detailed previously, these models predict the next token based upon the preceding sequence, utilising self-attention mechanisms to weigh the importance of different elements within the context. CoT prompting guides this generative process: instead of predicting the final answer directly, the model is steered towards generating tokens which represent intermediate steps in a reasoning sequence. Each generated step is then appended to the context. Consequently, the model's attention mechanism considers this augmented context – including the steps generated thus far – when predicting the subsequent token. This iterative process allows the model to condition later parts of its output on earlier parts, effectively employing the generated text as an explicit 'scratchpad' to manage the simulated reasoning process. It must be acknowledged that this process constitutes a simulation of reasoning, driven by the model learning statistical patterns from vast quantities of text data that include examples of arguments and explanations. LLMs do not manipulate internal symbolic logic representations, nor do they possess genuine comprehension in the human sense. Nevertheless, this simulation can be highly effective in practice. LLMs utilising CoT achieve strong performance on various reasoning and problem-solving benchmarks (e.g., arithmetic reasoning, commonsense reasoning); their simulated reasoning is often sufficient to solve complex tasks correctly. The key point for philosophical understanding, then, is that the outputs generated via CoT – these simulated arguments – are precisely the type of structured text that facilitates the mapping of dependence relations central to the account of understanding adopted here. These generated arguments can be complex and detailed, providing substantive material for philosophical analysis [Placeholder for examples of complex CoT philosophical arguments]. Moreover, the utility of these outputs parallels our engagement with human-generated philosophical texts; an argument need not be perfectly sound, nor entirely accepted by the reader, to enhance their understanding. Evaluating an argument, even one containing flawed steps, prompts the reader to critically examine the purported dependencies and refine their own representation of the conceptual network. This process of refinement seems equally applicable when the text being evaluated is a CoT output from an LLM. The effectiveness of this simulation is further supported by the observation that CoT capabilities emerge strongly with increased model scale (parameters numbering in the hundreds of billions or trillions) and are enhanced by refinement techniques such as Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which serve to improve the coherence, reliability, and instruction-following behaviour of the generated reasoning. Therefore, CoT connects directly to established philosophical methodology. It facilitates the generation of outputs structured like arguments, mirroring the sequential steps of premise-setting, inference, and conclusion characteristic of philosophical discourse. This structure, providing a potential pathway rather than just an endpoint, lends itself to analysis. It offers an artefact for philosophical scrutiny whose value derives from its effectiveness as a simulation, irrespective of its origin in a non-understanding source. The crucial outcome of these developments is that LLMs can produce highly effective simulations of reasoning. While they lack genuine understanding, their ability to generate step-by-step outputs significantly improves their performance on complex tasks. Models utilising CoT or similar structured reasoning approaches demonstrate strong results on various benchmarks requiring logical inference, arithmetic calculation, or commonsense reasoning. For many problem-solving purposes, this simulated reasoning proves remarkably capable, frequently arriving at correct solutions. Understanding how this simulation works is key. As established, LLMs operate by predicting the next token based on preceding context, using the Transformer architecture's self-attention mechanism to weigh the relevance of earlier tokens. They do not possess internal logic modules or conceptual understanding. Their ability to generate coherent reasoning chains stems from learning statistical patterns in their vast training data, which contains countless examples of human arguments, explanations, and step-by-step procedures. When prompted appropriately, the model activates these learned patterns, generating sequences that mimic the form of reasoning. The CoT process leverages the model's architecture effectively. By prompting the model to generate intermediate steps, these steps are added to the context window. The self-attention mechanism can then 'refer back' to these explicit steps when generating subsequent tokens, maintaining coherence and allowing later steps to build upon earlier ones. This turns the context window into a kind of 'scratchpad', enabling the simulation of a structured thought process. The quality and reliability of this simulation are further enhanced by the sheer scale of modern models and refinement techniques like Instruction Fine-Tuning (IFT) and Reinforcement Learning from Human Feedback (RLHF), which train the model to better follow instructions and produce outputs aligned with human judgments of coherence and accuracy. [^1]: There are of course degrees of significance, some philosophical works have a profound impact on the philosophical understanding of large swathes of people (Gettier, Kant etc.). [^2]: mention butlin, buckner and m etc. [^3]: Mention also the way they are trained being very different to how human's learn. ** # section scrap ## The Prospect of Significant Philosophical Generation A more interesting question, of course, is the extent to which LLMs can produce _novel_ outputs which significantly enhance their users philosophical understanding. Novel and significant are deliberately vague terms here but the idea here is simply to capture the range of impact a particular piece of philosophy might have: from the meaning of life etc. to a thought provoking article in a journal.[^1] The idea that their outputs can increase understanding is likely to face resistance. An initial objection one might have is that these systems do not themselves understand anything at all. They are stochastic parrots, incapable of doing anything other than arranging words in statistically plausible patterns. This may well me true: despite being based on neural networks the architecture of llms is a long way from that of a human mind[^2] [^3]; however, we should not assume that they need to reason in order to produce philosophically illuminating outputs. Butlin and Viebahn suggest that fine-tuned LLMs may produce outputs with a descriptive function, defined as "the function of conveying information to an observer or consumer system, so as to cause the consumer to behave as though some condition holds" (p. 4). While they argue this alone does not suffice for genuine assertion—since current LLMs cannot be sanctioned—they nonetheless highlight the significance of this capacity. Without fine-tuning, LLMs merely predict statistically likely sequences of words. Fine-tuning, however, involves additional training explicitly aimed at producing outputs that are accurate, relevant, and informative. By selectively rewarding outputs consistent with external sources or positively evaluated by humans, fine-tuning enhances the reliability of the information communicated. The significance of acquiring a descriptive function through fine-tuning lies in the generation of outputs reliable enough to aid users in refining their representation of dependence relations. In short, fine-tuning equips LLM outputs with a descriptive function, analogous to the way a thermometer is designed to indicate temperature. It might be then that fine-tuned LLMs should be given a hearing philosophically, indeed about anything else. Even if the system itself lacks understanding, and even it is not as consistently accurate as a thermometer, it has been made to provide correct answers. Correct answers are the sort of thing that would be useful for making one’s representation of dependency relations more accurate and more comprehensive. # scrap from 1000 words A similar low-level skill LLMs possess is restructuring text and the ideas within it, encompassing tasks such as summarization, format conversion, simplification, and expansion, alongside generating analogies or metaphors. This capability stems from their training on extensive datasets containing diverse examples of how information can be structured; the models identify statistical patterns in these structures and can generate text conforming to them. an LLM can usually provide correct translations of philosophical phrases – such as 'Cogito ergo sum' – or give answers to factual questions relevant to the field. Getting such information easily can arguably count as a minimal form of improved philosophical understanding. First, the rapid development of Large Language Models (LLMs) prompts a question about their potential role in philosophical work: To what extent might these systems contribute to philosophical understanding? A first look suggests that LLMs already offer contributions, though perhaps at a fairly basic level. It seems reasonable to say that these models can improve a user's understanding in certain simple ways; for example, However, one doubt about even this minimal contribution relates to the known tendency of LLMs to produce factually incorrect outputs, sometimes called 'hallucinations'. While this issue has been noted, its importance for the simplest kinds of query seems small. For simple translations or common-sense facts, the chance of current models giving wrong information seems quite low. Although such errors might appear with more complex data, like historical dates or who said what, their reliability for basic questions seems good enough to establish that LLMs can, and perhaps already do, contribute positively – at least in this small way – to philosophical understanding. Still, their potential reach arguably goes beyond these basic uses, suggesting a range of possible contributions. # Philosophical Understanding via Dependency Relations The account of philosophical progress presented by Dellsén et al. (2024), termed 'Enabling Noeticism', centres on a specific explication of understanding. Philosophical progress is seen as putting people in a position to increase their understanding of phenomena, where understanding itself consists in accurately and comprehensively representing the network of dependence relations surrounding a phenomenon. This involves mapping how the phenomenon depends on others, how others depend on it, and crucially, what it does _not_ depend on. This notion of understanding has two defining characteristics: it is 'epistemically undemanding' but 'robustly factive'. ## Key Characteristics of Understanding The 'epistemically undemanding' nature means that possessing understanding, in this specific sense, does not require the agent to meet traditional epistemic criteria like having justification, holding an outright belief, or possessing knowledge (in the standard epistemological sense) regarding the represented dependencies. For example, a physics student might accurately represent the dependencies described by quantum mechanics learned from a textbook, thereby gaining understanding in this sense, without personally possessing the deep empirical justification for the theory or even firmly believing in its literal truth. Similarly, grasping a philosophical distinction (like act/omission) involves representing the different dependencies it posits, which counts as understanding even if one lacks justification for believing the distinction tracks a genuine moral difference. The rationale is that justification and belief are often instrumental in _leading_ us to adopt accurate representations, but are not constitutive of the understanding conferred by the representation itself. Conversely, the 'robustly factive' nature means that the degree of understanding is determined by how well the representation corresponds to the _actual_ network of dependence relations in the world. Accuracy (correctness of the represented dependencies) and comprehensiveness (completeness in capturing the relevant network) are the measures. A representation must map onto reality to confer understanding. For instance, the shift from the inaccurate dependencies of Ptolemaic astronomy to the more accurate dependencies of Copernican astronomy constituted an increase in understanding because the representation better reflected the actual structure of the solar system. Likewise, if consequentialism accurately describes the dependencies determining moral rightness (i.e., solely consequences), then representing these dependencies confers more understanding than representing the dependencies posited by a factually incorrect deontological theory, regardless of the latter's coherence or intuitive appeal. Understanding, on this view, is about getting the dependencies right. ## Increasing Understanding: Specific Cognitive Changes Increasing philosophical understanding, therefore, involves cognitive changes that result in a more accurate or comprehensive representation of the relevant dependency network. Specific examples include: 1. **Replacing a Representation with a More Accurate One:** Discarding an existing representation of dependencies (e.g., personal identity depends on bodily continuity) and adopting a new one (e.g., personal identity depends on psychological continuity) after processing arguments (like Parfit's) suggesting the new representation corresponds more closely to the actual dependencies governing identity. 2. **Modifying a Representation due to Falsifying Information:** Processing information (e.g., Gettier cases) that reveals an inaccuracy in one's current representation ('Knowledge depends only on JTB') leads to modifying it (e.g., 'Knowledge depends on JTB + an anti-luck condition') to better align with the actual, more complex dependencies. 3. **Refining Representational Categories and Specifying Dependencies:** Processing a distinction (e.g., patterned vs. historical principles of justice) leads to replacing a general representation ('Justice depends on equality') with a more precise one specifying different dependencies for refined categories ('Justice in holdings depends on historical acquisition/transfer, not on patterns like equality'). 4. **Adding Dependency Relations to a Representation (Increasing Comprehensiveness):** Analysing an argument (e.g., Descartes' Cogito) reveals an implicit dependency (on 'Whatever thinks, exists'). Updating the representation to include this dependency makes the map of the argument's inferential structure more complete and accurate. 5. **Constructing a Meta-Representation of Dependencies Between Representations:** Mapping the relationships between different theories (e.g., Type-Identity Physicalism vs. Functionalism) by representing how one depends on rejecting premises of the other, or how they posit dependencies differently, increases comprehensive understanding of the theoretical landscape itself. 6. **Specifying Conditions on Dependency Relations:** Refining a representation ('Aesthetic value depends on formal properties') by adding conditions specifying _when_ this dependency holds versus when other dependencies (e.g., on context or intention) are operative, based on analysing different types of cases (e.g., abstract vs. conceptual art). ## Mechanisms for Enhancing Representations The primary mechanisms driving these cognitive changes appear to be twofold: - **Exposure to New Information:** This provides the essential input for forming, correcting, adding to, or replacing representations. Sources include texts, arguments, empirical data, counterexamples, and distinctions encountered through various media. - **Thinking/Analysis/Reflection:** This involves the active processing required to structure, integrate, evaluate, and refine representations based on information. Activities include logical analysis, comparison of theories, synthesis of ideas, identifying implicit structures, and applying concepts to cases. These two mechanisms are not really separable; thinking operates on information, and information prompts thinking. # another old draft: 19 Apr 2025 ## Introduction The advancement of artificial intelligence, particularly the development of Large Language Models (LLMs) like GPT-4, raises questions concerning their potential role in domains traditionally considered human, including philosophical inquiry. These models demonstrate proficiency in generating fluent, coherent, and contextually relevant text; this prompts discussions about their capacities for understanding, reasoning, and potentially creativity. This paper explores a specific aspect of this topic: the potential for LLMs to enhance human philosophical understanding. The central thesis proposed here is that LLMs can serve as effective tools to aid humans in achieving a specific type of philosophical understanding – namely, the accurate and comprehensive representation of dependence networks, following the account developed by Dellsén et al. (2024) – without needing to possess such understanding themselves. Consequently, this analysis distinguishes between the functional utility of LLMs as aids to human cognition and the separate question of whether these systems possess understanding, beliefs, or semantic competence in the human sense. The focus here is on the former: how might these pattern-matching systems contribute to the process of philosophical reflection and the refinement of understanding in human users? To address this, the paper will first outline the target conception of understanding, justifying its use for this specific inquiry. It will then introduce LLMs and their relevant capabilities, clarifying the scope of the inquiry regarding LLM cognition, before explaining in detail the mechanisms by which they might facilitate the representation of dependence networks. Finally, it offers concluding thoughts on the implications of using these tools in philosophy. ## 1. Understanding as Representing Dependence Networks To assess how LLMs might enhance philosophical understanding, we first require a definition of this understanding. This paper adopts the framework proposed by Dellsén et al. (2024). Their approach, which explains understanding in terms of representing dependence networks, can be situated within a broadly Sellarsian tradition aiming to understand 'how things in the broadest possible sense of the term hang together in the broadest possible sense of the term' (Sellars 1962, p. 1). Dellsén et al. offer a specific way to construe this 'hanging together' for the purposes of evaluating philosophical progress, focusing on the representation of dependence relations. Their account differs from analyses focused solely on the ordinary meaning of 'understanding'; its stated aim is instead "to explicate it in a way that makes it suitable for being deployed in an account of philosophical progress" (Dellsén et al., 2024, p. 674). The central claim holds that the degree to which a subject understands a phenomenon corresponds to the accuracy and comprehensiveness of their representation of the network of dependence relations relevant to that phenomenon. #### Interlude with a bit more detail on Sellers >The aim of philosophy, abstractly formulated, is to understand how things in the broadest possible sense of the term hang together in the broadest possible sense of the term. This statement encapsulates Sellars's conception of philosophy as fundamentally concerned with achieving a **synoptic vision**. It's not merely about analyzing individual concepts or arguments in isolation, but about constructing a coherent and comprehensive understanding of reality in its entirety. Let's unpack the key components: 1. **"Things in the broadest possible sense of the term"**: This signals an extremely wide **ontological scope**. Sellars isn't just talking about physical objects. "Things" here encompass entities, concepts, events, properties, norms, values, possibilities, theoretical posits, linguistic items, mental states, social practices – essentially, anything that can be thought about or referred to. This includes, crucially, the items populating both common-sense frameworks and scientific theories. 2. **"Hang together"**: This metaphor signifies **integration, coherence, and systematic interconnection**. Philosophy's task is to understand the relationships _between_ these diverse "things." How do they relate? Causally? Logically? Conceptually? Normatively? Explanatorily? Constitutively? Sellars is interested in mapping out these intricate connections to form a unified picture. It implies seeking consistency and resolving apparent conflicts between different domains of discourse or frameworks of understanding. 3. **"In the broadest possible sense of the term" (second occurrence)**: This reinforces the comprehensive nature of the **modes of integration** sought. Philosophy shouldn't restrict itself to only one type of connection (e.g., only causal or only logical relations). It must consider _all_ the ways different aspects of reality and thought interrelate to form a coherent whole. **Relation to Sellars's Project:** This conception directly motivates Sellars's central project in PSIM: the confrontation and reconciliation of the **Manifest Image** and the **Scientific Image**. • The **Manifest Image (MI)** is the framework in which humans understand themselves as persons with thoughts, intentions, and norms, existing within a world of perceptible objects and qualities. It's the refined image of common sense, incorporating empirical and conceptual sophistication but fundamentally centered on the category of the person. • The **Scientific Image (SI)** is the framework derived from the postulational methods of the sciences, positing unobservable entities (particles, fields, etc.) to provide a comprehensive causal-explanatory account of the world, including humans as complex physical systems. The philosophical task, as Sellars sees it, is precisely to understand how these two powerful and seemingly conflicting images "hang together." Can the world described by the SI accommodate the persons, norms, and meanings central to the MI? The SI appears to claim completeness in the descriptive/explanatory order, potentially threatening the ontological standing of entities unique to the MI. Sellars's goal isn't simple reduction or elimination but a **"stereoscopic fusion"**. He aims to show how the indispensable conceptual framework of the MI (especially concerning normativity, personhood, and meaning) can be coherently integrated with the descriptive and explanatory power of the SI, achieving a single, unified vision of reality. The famous quote thus sets the stage for this complex integrative project, defining philosophy's core ambition as the pursuit of ultimate conceptual coherence across all domains of thought and reality. These dependence relations are conceived as objective, worldly relations that often underpin explanatory practices. As Dellsén et al. describe, using Kim's (1994) work as a basis: The idea that understanding is centrally concerned with representing dependence relations dates back to Kim (1994). For Kim, dependence relations are the ontological correlates of explanation – they are the worldly relations that make it true that something explains something else. In much of empirical science, the paradigmatic dependence relation is causation... It is a matter of contention which other relations are genuine dependence relations, but they may include constitution, grounding, mereological dependence, truthmaking, conceptual containment, and/or supervenience. (Dellsén et al., 2024, p. 675) The framework does not commit to a precise list of genuine dependence relations, assuming that understanding involves representing whichever relations actually obtain in the domain under investigation. Furthermore, understanding a phenomenon X requires grasping its position within a wider network representation; it extends beyond merely knowing the factors upon which X depends. [...] understanding X is a matter of representing both how X depends on various other phenomena, and how further phenomena depend on X itself. In other words, the extent to which one understands X is a matter of how one represents the network of dependence relations running both to, and from, X. (Dellsén et al., 2024, p. 675) Such a representation includes facts about both existing dependencies (positive dependencies) and dependencies that do not exist (negative dependencies, for instance, that X lacks dependence on Y). Understanding, on this view, admits of degrees. Its extent relates to two criteria: accuracy and comprehensiveness. Accuracy refers to how correctly the representation maps the dependencies and non-dependencies that actually exist. Comprehensiveness refers to the extent to which the representation includes all the relevant phenomena and the dependence relations (or lack thereof) connecting them. These criteria can sometimes be in tension. Such tension potentially leads to trade-offs like idealisation (sacrificing some accuracy for broader scope, e.g., simplifying the conditions for causation to map a wider explanatory network) or abstraction (sacrificing some comprehensiveness for greater precision, e.g., focusing solely on the logical dependencies within a formal system) (Dellsén et al., 2024, p. 675). A feature of this account, relevant here, is that it is "robustly factive," requiring the representation to correspond to the actual facts about dependencies, yet also "epistemically undemanding" (Dellsén et al., 2024, p. 674, 676). Understanding X, in this specific sense, does not require having justification for, or even belief in, the propositions represented within the dependence network. [...] on the explication of understanding with which we will operate, understanding X does not imply having the type of epistemic justification that is required for knowing any particular proposition about X. [...] More importantly for our purposes, an explication of understanding that does not imply justification is arguably better suited for spelling out a plausible understanding-based account of philosophical progress. (Dellsén et al., 2024, p. 676) Thus, understanding in this specific sense does not require justification or belief regarding the represented dependencies. This notion—understanding as the accurate and comprehensive representation of networks of dependence relations, without requiring justification—guides the subsequent analysis. One might object that this definition is too thin, perhaps omitting aspects like evaluative judgment or practical wisdom often associated with deep philosophical understanding. While acknowledging that other, richer notions of understanding exist, the Dellsén et al. account is adopted here precisely because its focus on representing relational structures provides a clear, measurable target against which the potential contribution of information-processing tools like LLMs can be assessed. Enhancing philosophical understanding, therefore, involves refining this mental representation of the dependence network. When encountering philosophical arguments, thought experiments, or new information, an individual engages in cognitive processing to evaluate whether incorporating this new input would yield a representation that is more accurate or more comprehensive. Consider wanting to enhance one's understanding of the nature of time. Reading debates on A-series versus B-series theories might refine an initially simplistic map. An inaccurate link between temporal passage and mere change might become more accurate by distinguishing subjective passage (perhaps dependent on A-series views) from objective temporal ordering (central to B-series views). Comprehensiveness might increase by incorporating dependencies linking temporal theories to broader metaphysical positions (e.g., presentism, eternalism) or findings in physics, and also mapping dependencies from theories of time to concepts like agency or causation. Clarity improves by explicitly mapping negative dependencies (e.g., B-series ordering not depending on subjective experience) or specifying the type of dependence (e.g., conceptual, metaphysical, grounding). Based on this evaluation, the individual adjusts their mental model – adding, removing, or altering represented dependencies to better reflect the actual relations. Consistent with the epistemically undemanding nature of the account, the resulting revised representation constitutes enhanced understanding simply by virtue of its improved accuracy or comprehensiveness, even without accompanying justification or belief regarding every element. Having defined the target concept of understanding relevant to this paper – the accurate and comprehensive representation of dependence networks, achieved through a process of refining one's mental map – we can now turn to the central question. The subsequent investigation addresses how Large Language Models might function as tools to assist humans in this specific process of refining their understanding. ## 2. LLMs Large Language Models (LLMs) are computational systems based on large artificial neural networks, trained on massive datasets of text with the core task of predicting the next word (or token) in a sequence, given the sequence of words that came before. To perform this predictive task well across diverse contexts, the model must learn the underlying statistical patterns and regularities of language. LLMs typically employ the Transformer architecture (Vaswani et al. 2017), using self-attention mechanisms to dynamically weigh the importance of different words in the prior context, enabling them to handle the long-range connections necessary for coherent text generation and understanding. Given the vast scale of their training data compared to their fixed number of internal parameters, LLMs cannot simply memorise the input text. Effective prediction necessitates compression: the model must identify and efficiently represent recurring patterns and structures within the language data. This process forces the model to develop internal states or encodings – often referred to as _latent representations_ – that capture the statistical essence of the language it has processed, including grammar, common phrasing, and semantic associations. These latent representations are physically realised across the different components of the transformer architecture. The embedding layer learns vector representations for words or tokens where geometric relationships (like distance or direction) reflect semantic similarities and usage patterns observed in the training data. Within the network's layers, self-attention heads learn functions for passing information between token representations, often specialising such that some heads track local syntactic agreement while others manage longer-range thematic or referential links. The feed-forward (MLP) layers further process these representations, and through training, parts of these layers or specific patterns of activation within them may become sensitive to more abstract linguistic or conceptual features. These representations are typically distributed, meaning specific information is often encoded across many parameters rather than being localised in one place (a phenomenon sometimes referred to as superposition). For contemporary LLMs, the training process does not end with pre-training on raw text. An additional stage of fine-tuning, often involving Reinforcement Learning from Human Feedback (RLHF) (Christiano et al. 2017; OpenAI 2022), is commonly applied. In RLHF, human reviewers provide feedback (e.g., by ranking different model outputs), which is used to train a 'reward model' that predicts human preferences. The LLM is then further trained to maximise the scores assigned by this reward model, thereby aligning its behaviour more closely with desired qualities like helpfulness, honesty, and harmlessness. This alignment process, as argued by Butlin and Viebahn (2024), introduces criteria related to accuracy and reliability, giving the model's outputs "descriptive functions" – a functional orientation towards providing informative and correct responses, beyond simply predicting statistically likely sequences. The result of these training processes – pre-training focused on prediction, and fine-tuning focused on alignment – is that LLMs develop internal latent representations. These learned latent representations inevitably come to encode information reflecting the relational structures discussed in the training data, corresponding to how terms related to causation, conditionality, constitution, logical entailment, grounding, and other forms of dependence are typically used and interconnected in language. Connecting this back to the framework outlined in Section 1, the internal network of refined latent representations within an LLM can be understood as constituting a vast repository of _candidate_ dependence relations. Because these representations are learned from text that often reflects real conceptual, logical, or worldly dependencies, and because they are oriented towards generating informative outputs (descriptive representations), the text generated by LLMs can serve as valuable input for a philosopher engaged in the task of mapping dependence networks (à la Dellsén et al.). The model's output, reflecting the dependencies captured in its latent representations, provides structured information that the philosopher can evaluate. Thus, the LLM functions as a tool, providing access to representations of dependencies without needing to possess understanding itself. Acknowledging potential limitations, such as biases inherited from training data or distortions introduced by the alignment process, this capacity forms the basis for exploring the practical applications of LLMs in philosophical inquiry in the subsequent sections # Part Two: Practical Applications for Motivated Philosophical Prompters - In the following section I would like to make some suggestions as to how the ideas I have just outlined can be put into practice. Before that I want to narrow my focus slightly so as to avoid the following issue. - It is all very well to try to make the sorts of connections that I just did, but day to day encounters with chat gpt written student papers students who might charitably be said to be 'using ai' philosophy does not brought a smirk to many readers' faces; indeed, it seems likely that many readers' primary encounters with 21st-century artificial intelligence will take the form of deeply banal, entirely worthless student essays generated by ChatGPT, featuring such banalities as 'let's delve into this interesting topic'. Beyond this superficial usage, specific failure modes pertinent to philosophical methodology may arise even in more complex interactions, potentially hindering rather than enhancing understanding. Cognitive outsourcing presents one such risk: if a user primarily relies on the LLM to structure arguments or synthesise information, they might bypass the internal cognitive work—analysis, synthesis, critical reflection—needed to develop their own robust representation of the dependence network (as defined in Part 1). Difficulties in evaluation pose another challenge. A user without sufficient background knowledge might struggle to assess LLM output effectively for subtle errors or inaccuracies, potentially incorporating flawed dependency claims into their representation. This points towards a potential evaluation impasse, where those most in need of accurate information might be least equipped to verify it. Acknowledging the substance of these concerns regarding misuse and certain interaction dynamics, this investigation nevertheless sets them aside to focus on a different question: Can a 'motivated philosopher prompter' use LLMs effectively to enhance their understanding? For the purposes of this inquiry, such a prompter is defined as an individual, often possessing philosophical training, who engages critically and interactively with an LLM. Such an individual's goal is assumed not to be replacing their own thinking or generating a final product automatically. Instead, the aim is to employ the LLM strategically as a tool—perhaps as an assistant for brainstorming, exploring conceptual connections, structuring arguments, or accessing information—with the aim specifically of enhancing their _own_ understanding, construed here as making their internal representation of philosophical dependence relations more accurate and comprehensive (cf. Section 1). _(Drafting paused here as requested)_ ## 4. Minimal Understanding Enhancements: Refining Dependence Networks The preceding discussion outlined how Large Language Models might, in principle, facilitate the representation of dependence networks characteristic of philosophical understanding. To make this potential more concrete, it is helpful to consider how such enhancements operate at the level of individual cognition, focusing on what might be termed 'minimal understanding enhancements'. These are the incremental refinements – often small adjustments in accuracy or minor expansions in scope – that occur when an individual encounters and processes new philosophical information or argumentation. While philosophical progress often involves major shifts in perspective, much of the day-to-day cognitive work involves these smaller steps, gradually improving the fidelity and coverage of one's internal representation of the relevant dependence structures. According to the framework adopted here (Dellsén et al. 2024), each instance where a represented dependence relation becomes more accurate (e.g., correcting a misidentified dependency, specifying the type of dependence more precisely) or where the represented network becomes more comprehensive (e.g., incorporating a previously overlooked relevant factor or mapping a new connection) constitutes an increase in understanding, however slight. These minimal enhancements are the building blocks of more substantial gains in philosophical insight. They represent the micro-level adjustments through which a motivated individual progressively refines their mental map of how things hang together within a specific philosophical domain. Consider, for instance, an individual reflecting on the nature of memory and its relation to past events. Initially, their representation might be relatively simple: they might represent a current memory state (M) as depending directly on a past event (E). This could be depicted crudely as: [Past Event E] ---> [Current Memory M] Here, the arrow represents a general dependence relation, perhaps loosely conceived as causal or informational. While not entirely wrong, this representation lacks detail; it is insufficiently comprehensive and potentially inaccurate by conflating different phenomena under the single node 'M'. Now, suppose this individual encounters philosophical discussions distinguishing between episodic memory (the ability to mentally re-experience a past event from a first-person perspective, often associated with autonoetic consciousness) and semantic memory (knowledge of facts about the past, lacking the experiential quality of reliving, associated with noetic consciousness) – a distinction prominent in both psychology (Tulving 1985) and philosophy of mind. Processing this distinction prompts a refinement of their dependence network representation. The single node 'M' might be replaced by two distinct nodes: 'Episodic Memory of E' (M<0xE2><0x82><0x91><0xE1><0xB5><0x96><0xE1><0xB5><0xA2>) and 'Semantic Memory that E occurred' (M<0xE2><0x82><0xA2><0xE1><0xB5><0x89><0xE1><0xB5><0x90>). The dependence relations can now be specified more accurately. Episodic memory might be represented as depending not just on the event E itself, but perhaps on the original first-person experience of E, and potentially on specific cognitive capacities related to self-awareness and mental time travel. Semantic memory, conversely, might depend on receiving information about E (which could originate from experiencing E, but also from testimony, inference, etc.) and capacities for factual recall. The revised network might include representations like: [First-person Experience of E] ---> [M<0xE2><0x82><0x91><0xE1><0xB5><0x96><0xE1><0xB5><0xA2>] _(Dependence: Grounding/Causal basis)_ [Information that E occurred] ---> [M<0xE2><0x82><0xA2><0xE1><0xB5><0x89><0xE1><0xB5><0x90>] _(Dependence: Informational basis)_ Furthermore, the refinement might involve adding negative dependencies, such as representing that M<0xE2><0x82><0x91><0xE1><0xB5><0x96><0xE1><0xB5><0xA2> does _not_ solely depend on factual information transmission, or that M<0xE2><0x82><0xA2><0xE1><0xB5><0x89><0xE1><0xB5><0x90> does _not_ necessarily depend on a first-person perspective. This adjustment—replacing a single, underspecified dependency with a more articulated network differentiating types of memory and their distinct dependencies—constitutes a minimal understanding enhancement. The representation has become both more accurate (distinguishing genuinely different relations and relata) and more comprehensive (incorporating finer-grained concepts). This example illustrates the type of cognitive refinement central to the paper's thesis. The process involves modifying one's internal map of dependencies based on new conceptual distinctions or arguments. It is precisely this process that LLMs, used appropriately by a motivated prompter, can potentially facilitate. An LLM could, for instance, be prompted with basic questions about memory, leading to outputs that introduce or clarify the episodic/semantic distinction. More specific prompts could yield targeted information: - A prompt like "Explain the distinction between episodic and semantic memory" might elicit a response defining episodic memory in terms of subjective reliving ('mental time travel' into the past, involving autonoetic consciousness) and semantic memory as context-free factual knowledge (noetic consciousness), perhaps citing Tulving. This directly helps the user differentiate the nodes (M<0xE2><0x82><0x91><0xE1><0xB5><0x96><0xE1><0xB5><0xA2> vs M<0xE2><0x82><0xA2><0xE1><0xB5><0x89><0xE1><0xB5><0x90>) and associate them with distinct properties, enhancing representational accuracy. - Asking "What cognitive capacities does episodic memory depend on?" could generate a list including pattern completion, scene construction, self-awareness, and the ability to bind contextual details (time, place, perceptual information) to an event representation. This output helps the user map the _inputs_ or grounding conditions for M<0xE2><0x82><0x91><0xE1><0xB5><0x96><0xE1><0xB5><0xA2>, making the network more comprehensive and specifying the nature of the dependencies (e.g., enabling conditions, constitutive elements). - A request to "Summarise arguments linking episodic memory to imagining the future" might produce text outlining the constructive episodic simulation hypothesis (Schacter & Addis 2007), suggesting that the same neural and cognitive mechanisms underlie both remembering the past and imagining novel future events. This helps the user map _outputs_ or consequences of M<0xE2><0x82><0x91><0xE1><0xB5><0x96><0xE1><0xB5><0xA2>, specifically its dependence relationship with prospective thought, again increasing comprehensiveness. - Prompting "Are episodic and semantic memory fully independent?" could lead to a discussion of interactionist views, such as how semantic knowledge structures provide schemas for encoding and retrieving episodic memories, or how repeated episodic experiences can consolidate into semantic knowledge. This encourages the user to refine the relationship _between_ M<0xE2><0x82><0x91><0xE1><0xB5><0x96><0xE1><0xB5><0xA2> and M<0xE2><0x82><0xA2><0xE1><0xB5><0x89><0xE1><0xB5><0x90>, potentially adding bidirectional dependencies or qualifying their distinctness, enhancing accuracy. By generating text that articulates these definitions, dependencies, arguments, and nuances (drawing on its inferential semantic competence and vast training data, as discussed in Section 3), the LLM provides the user with structured information directly relevant to refining their dependence network. The LLM does not possess the understanding itself, nor does it perform the cognitive act of refinement. Rather, its output serves as organised, relevant input that the user critically evaluates and integrates, facilitating their own cognitive process of achieving minimal (and potentially cumulative) understanding enhancements. The tool assists the user in identifying nodes more clearly, specifying relations more accurately, mapping connections more comprehensively, and ultimately increasing the overall fidelity of their mental map, aligning with the definition of understanding adopted herein. #### Enhancement through Linguistic Refinement A distinct but related way LLMs facilitate minimal understanding enhancements is through refining the language used to articulate philosophical ideas, which in turn sharpens the representation of the dependence network. Consider the common scenario where a philosopher attempts to formulate a precise claim or definition and uses an LLM as a writing assistant, perhaps asking "How can I say this more accurately?" or querying for alternative phrasings. The LLM, drawing on its vast knowledge of semantic relationships and contextual nuance, might suggest synonyms or reformulations that better capture the intended meaning. This is not merely a stylistic improvement; it directly impacts the representation of dependencies. For example, if one initially represents the connection between consciousness and brain states using the general term 'depends on', asking an LLM for more precise alternatives in this context might yield terms like 'is constituted by', 'supervenes on', 'is realised by', or 'emerges from'. Each of these terms corresponds to a different, specific type of dependence relation debated in philosophy of mind. Evaluating the LLM's suggestions forces the user to clarify their intended claim, selecting the term that most accurately reflects the specific dependency they aim to represent in their mental model. Choosing 'supervenes on' over the vague 'depends on' or the stronger 'is constituted by' directly increases the _accuracy_ of the represented dependence relation within the user's network. This process of LLM-assisted linguistic refinement thus translates directly into conceptual clarification and a more accurate mapping of the dependence network structure. Even seemingly mundane requests for synonyms or clearer phrasing can, therefore, function as prompts for minimal enhancements in understanding, improving the fidelity of the user's representation by forcing greater precision in articulating the dependencies involved. Small improvements in how we state ideas might seem minor compared to major conceptual shifts. Yet using language clearly is important for philosophical work, as it helps improve the accuracy of our dependency representations. Suppose a philosopher exploring the link between consciousness and brain states initially uses the broad term 'depends on'. Asking an LLM for more precise alternatives could yield 'supervenes on' or 'is constituted by'. Choosing between these requires the philosopher to clarify the exact relationship being represented, making their dependency map more accurate. This clarification then allows for more focused questions: having specified 'supervenience', one can investigate its features or consequences. In this way, understanding builds step-by-step, creating a ratchet effect where each clarification supports the next stage of thought. This process is similar to the careful conceptual groundwork seen in thinkers like Kant and Frege. LLMs can help this clarifying process by suggesting better words or phrasing. While these minimal enhancements differ from the large-scale breakthroughs discussed next, they help build the foundation needed for deeper understanding. ## 5. Comprehensive Breakthroughs The focus thus far has been on 'minimal understanding enhancements' – the gradual, incremental refinements of accuracy and comprehensiveness in an individual's representation of dependence networks. However, the history of philosophy, science, and other intellectual domains, as well as individual cognitive life, also features episodes of more dramatic change: 'comprehensive breakthroughs'. These involve relatively rapid, large-scale restructuring or updating of dependence network representations, potentially leading to significant shifts in perspective or paradigm. Such breakthroughs might occur at the personal level, either within a specific intellectual domain (e.g., grasping a revolutionary scientific theory or a complex philosophical system) or more holistically (akin to L.A. Paul's (2014) notion of transformative experiences that reshape one's entire value structure and outlook). They can also occur at the collective level, such as Kuhnian paradigm shifts within scientific communities (Kuhn 1962), where the commonly accepted network of concepts and dependencies undergoes radical revision. Given the argument that LLMs primarily function by processing linguistic patterns and generating text based on learned statistical relationships, a crucial question arises: Can these tools contribute to facilitating or catalysing such comprehensive breakthroughs, or is their role necessarily confined to the more modest, minimal enhancements discussed previously? Several potential limitations suggest caution. Breakthroughs often seem to involve deep conceptual creativity, a grasp of significance, and sometimes an 'Aha!' moment of insight – capacities linked to genuine understanding, intentionality, and subjective experience, which LLMs lack. Furthermore, many breakthroughs rely on empirical grounding or embodied interaction with the world, areas where LLMs are inherently limited. However, these objections may not be decisive. Regarding the lack of genuine understanding and goals, the argument developed earlier holds: if LLMs possess a descriptive or representational function enabling their outputs to accurately map dependence relations, this functional capacity might be sufficient to provide crucial input for a human user on the verge of a breakthrough. The LLM need not _understand_ the significance of the connections it presents; it merely needs to present them in a way that allows the human user to grasp their significance. Similarly, the LLM does not need to experience an 'Aha!' moment; its function could be to generate the novel juxtaposition, synthesis, or representation of dependencies that _triggers_ such a moment in the user. The critical cognitive event remains human, but the LLM could provide the essential stimulus derived from its processing of vast information structures. The objection concerning creativity versus recombination raises deeper issues about the nature of breakthroughs themselves. If conceptual revolutions are not purely random events but involve discoverable, albeit perhaps complex, patterns of conceptual restructuring, then the question becomes whether these patterns are representable within the high-dimensional latent spaces learned by LLMs. It seems plausible that patterns underlying even radical shifts in dependence networks, insofar as they are expressible or derivable from the vast corpus of human text and data LLMs are trained on, could be captured. If so, an LLM might, in principle, generate outputs that reflect pathways towards these restructurings, perhaps by identifying deep structural analogies, revealing hidden assumptions, or synthesising information in radically novel ways. The burden would then fall on critics to articulate why the non-random patterns underlying breakthroughs should be systematically unrepresentable by such models. The limitation regarding grounding and embodiment remains significant, likely restricting the potential for LLM facilitation of breakthroughs in empirically grounded sciences or areas of philosophy deeply connected to lived experience. However, in more abstract domains – such as mathematics, formal logic, and potentially certain areas of metaphysics or epistemology – where breakthroughs may depend more heavily on manipulating complex symbolic systems and exploring abstract structures, the LLM's lack of embodiment may be less detrimental. In these areas, its ability to process information, identify patterns, explore logical consequences, and structure complexity at scale could prove particularly valuable. Therefore, while LLMs cannot _have_ breakthroughs, it remains plausible that they could function as powerful catalysts or accelerators in the process. By synthesising information, revealing latent structures, facilitating the exploration of conceptual spaces, and structuring complexity, they might significantly aid a prepared human mind in achieving the large-scale network restructuring characteristic of a comprehensive breakthrough, particularly within abstract domains. Their role may be less about generating the core insight ex nihilo and more about creating the conditions under which such insights are more likely to occur in the human user. # old section 2 and 3 18th April 2025 ## 2. Large Language Models and Their Potential as Philosophical Tools Recent developments in artificial intelligence, specifically the emergence of Large Language Models (LLMs), present possibilities concerning the enhancement of philosophical understanding as defined above. LLMs are computational systems built upon large artificial neural networks, most commonly the Transformer architecture (Vaswani et al. 2017), trained on extensive datasets comprising trillions of tokens of text and code. Their primary training objective typically involves predicting subsequent linguistic tokens in a sequence; this enables them to generate coherent and contextually relevant text in response to user prompts. A central mechanism underlying their capabilities is self-attention, which allows the model to dynamically weigh the significance of different parts of the input context when producing output, enabling the handling of long-range dependencies in text. Many of their behaviours, such as apparent reasoning or few-shot learning, are often considered emergent properties resulting from the scale of the models and the extent of their training data (Brown et al. 2020). While the basic function of predicting the next token is grounded in statistical patterns learned from the training corpus, this alone provides a basis for their utility in philosophy. By mastering the statistical regularities of language, LLMs become adept at reproducing complex linguistic structures, including the ways philosophers express arguments, definitions, objections, and dependencies. They learn the common linguistic structures of philosophical discourse found in their training data (which often includes academic articles, books, and encyclopedias). This capability allows them to generate text that articulates standard philosophical positions, outlines arguments, or explains concepts in ways that reflect the dependencies discussed within the philosophical literature they were trained on. The volume of text processed means they encounter countless instances expressing logical entailment, conceptual analysis, causal claims, and grounding assertions, learning the statistical contexts in which terms signifying these relations appear. This proficiency in reproducing complex linguistic structures, including philosophical arguments, forms the basis for their potential utility as tools. However, assessing this utility requires a clear methodological focus, given ongoing philosophical discussion regarding whether LLMs possess genuine understanding, communicative intentions, or semantic grounding comparable to humans (cf. Millière & Buckner 2024). Arguments questioning their cognitive status often highlight their lack of embodiment, perceptual experience, intrinsic goals, or stable beliefs, suggesting their linguistic abilities arise from pattern matching rather than human-like cognition (cf. Harnad 1990; Bender & Koller 2020). Acknowledging the substance and complexity of these debates, this paper sets them aside to concentrate specifically on the functional question: can LLMs, regardless of their intrinsic cognitive status, assist human users in refining their representations of dependence networks (as defined in Section 1)? The argument advanced here pertains solely to this functional capacity. This focus allows for an analysis of the demonstrable mechanisms by which LLM outputs might aid human cognition in this specific philosophical task, an analysis undertaken in Section 3. The utility of LLMs as tools for enhancing understanding is increased by processes applied after the initial pre-training phase. Pre-trained models are optimised solely for next-token prediction, irrespective of the truthfulness, usefulness, or safety of the generated text. To address this, models like ChatGPT undergo further stages of fine-tuning, often involving techniques like Reinforcement Learning from Human Feedback (RLHF) (Christiano et al. 2017; OpenAI 2022). In RLHF, human reviewers compare and rank different model outputs based on criteria such as helpfulness, honesty, and harmlessness. This comparison data is used to train a reward model, which essentially learns to predict which kinds of outputs human reviewers prefer. This reward model then guides the LLM during a reinforcement learning phase, adjusting the LLM's parameters to increase the probability of generating outputs that the reward model scores highly. This process repeatedly guides the model to produce text that better aligns with human preferences and normative goals, for instance providing accurate information or avoiding harmful content. This fine-tuning process is relevant here. As Butlin and Viebahn (2024) argue, such processes introduce criteria related to accuracy, verifiability, or groundedness that go beyond mere statistical likelihood. Fine-tuning might reward outputs judged 'correct' (perhaps because they align with information from reliable sources used by the human raters) or penalise those deemed 'incorrect' or 'ungrounded'. This deliberate selection pressure, mediated by the reward model trained on human judgments, modifies the system's optimisation criteria. The function governing the generation of outputs shifts partially from pure statistical prediction towards alignment with these normative criteria. Consequently, outputs are produced not merely as statistically probable sequences, but, at least partly, because they satisfy conditions related to accuracy or reliability according to the learned reward model. Butlin and Viebahn term this acquiring descriptive functions. Although this does not equate to genuine assertion or belief in the human sense, it suggests the outputs of fine-tuned LLMs are functionally aimed at being informative and accurate in a way that pre-trained models are not. This development indicates their potential as tools for enhancing understanding: outputs generated with descriptive functions are, in principle, more suitable as reliable input for a human user aiming to construct an accurate representation of dependence networks. The fine-tuning pushes the model's output distribution towards regions favoured by human evaluators seeking correctness, making the generated text potentially reflect actual dependencies more closely. ## 3. How LLM Capabilities Facilitate Representing Dependence Networks Given that LLMs, particularly fine-tuned ones, can generate contextually relevant text oriented towards informativeness, how specifically might their capabilities assist a human user in the task of representing dependence networks more accurately and comprehensively? Several interconnected features, stemming from their architecture and training, contribute to this potential. These features allow them to function as aids in mapping the relational structures inherent in philosophical discourse. These mechanisms operate primarily by processing and structuring information found in language, suited to the task of mapping conceptual, logical, and explanatory relations. The focus here is on the practical application of these capabilities for philosophical work. One strength is processing and reproducing linguistic patterns that encode dependencies. Having been trained on vast amounts of text, including philosophical works, LLMs learn the common linguistic structures used to express dependence relations – phrases like "X causes Y," "P implies Q," "A is a necessary condition for B," "G grounds F," or "S constitutes P." When prompted appropriately, they can generate text articulating standard arguments, definitions, or explanations by completing these learned patterns. Stating known or widely discussed dependency claims (e.g., the dependence of an action's moral status on its consequences in utilitarianism, or the dependence of knowledge on justification in traditional epistemology) helps a user map established theoretical dependencies within a specific philosophical domain. This capability relates to inferential semantic competence, as discussed by Millière and Buckner (2024) following Marconi (1997). While referential competence (connecting words robustly to the world) might be limited, LLMs seem to excel at the inferential aspect – the web of word-to-word, concept-to-concept relationships expressed in language. They learn, as Piantadosi and Hill (2022) suggest, the functional roles of concepts based on their distribution and relations to other concepts within the linguistic system. This involves identifying synonyms and antonyms, understanding hyponymy and hypernymy, recognising entailment relations between sentences, and providing definitions or paraphrases. Since philosophical understanding, particularly concerning conceptual analysis and logical argumentation, relies heavily on mapping these inferential and conceptual dependencies – tracing the logical consequences of definitions, identifying conceptual necessities, comparing the structure of different theories – the LLM's strength in this area is applicable to the user's task of refining their dependence network representation. For example, a philosopher analysing a complex text by Spinoza could use an LLM to identify all sentences asserting a substance-mode dependence, or to trace the definitional dependencies between key terms like 'attribute' and 'substance'. Similarly, when evaluating a Gettier case, an LLM could be prompted to list the premises and conclusion. This makes the inferential structure (and the dependence of the conclusion on the premises) explicit for analysis, potentially highlighting where a premise fails or the inference is invalid. This ability to extract and structure inferential relations from text aids mapping dependencies. Beyond reproducing explicit patterns, the internal representations learned by LLMs may also help discover or consider other dependencies. LLMs represent words and concepts as vectors in a high-dimensional space, where semantic relationships are often represented geometrically (Millière & Buckner 2024). Concepts represented by vectors close to each other in this space might share underlying features or have similar dependence relations (e.g., grounding, constitution). For example, terms referring to different kinds of normative reasons (moral, prudential, epistemic) might cluster in the vector space; their relative positions could potentially reflect structural similarities or differences in how these reasons are thought to ground obligations or beliefs. The model's ability to perform analogical reasoning (e.g., identifying A:B :: C:? relationships via vector arithmetic, Mikolov et al. 2013) might show parallel dependency structures across different philosophical problems or domains. For instance, based on textual patterns, it might suggest an analogy between the dependence of aesthetic value on formal properties in art and the dependence of semantic content on syntactic structure in language, suggesting the user consider a shared underlying principle of structural dependence. Also, as Andreas (2022) suggests, the process of efficiently compressing vast amounts of text might lead LLMs to implicitly develop internal representations that function as partial world models – capturing latent variables or structures that generate the observed text patterns. In philosophy, this could mean developing implicit models of abstract domains, such as ethical frameworks (e.g., representing the relationships between utility, rights, and duties within different consequentialist and deontological systems) or metaphysical systems (e.g., modelling the allowed relations between universals and particulars in different ontological schemes). If an LLM develops even a basic model of the entities and relations within a specific philosophical theory, its outputs when discussing that theory might reflect the dependencies in that model. For example, it might generate scenarios illustrating the consequences of adopting a particular stance on personal identity (e.g., psychological continuity vs. bodily continuity) that consistently reflect the dependencies implied by that stance, potentially revealing implications. This capacity to synthesise broad knowledge and potentially model underlying structures allows LLMs to generate lists of factors relevant to a phenomenon or suggest connections the user might not have considered, thereby increasing the comprehensiveness of the dependence network representation. The Transformer architecture itself provides features for handling the complexity of philosophical arguments. The self-attention mechanism (Vaswani et al. 2017) enables the model to track relationships between terms and concepts across extended passages. This helps accurately represent dependencies in arguments where conclusions rely on premises stated much earlier, or where dependencies involve multiple interrelated concepts introduced over time. For example, when analysing a lengthy text debating the dependence of moral responsibility on free will, the self-attention mechanism helps the model keep track of how 'free will' was defined or constrained in earlier sections when interpreting later sentences about praise, blame, or punishment. By maintaining contextual coherence, the generated text is more likely to correctly represent the intended dependencies relative to the specific context provided, increasing the accuracy of the user's mapping of that particular argumentative network (Millière & Buckner 2024). Additionally, prompting techniques like Chain-of-Thought (CoT) (Wei et al. 2022) can produce step-by-step reasoning from LLMs. This process states the intermediate inferential dependencies in an argument or explanation, making the path from premises to conclusion clearer. For a human user attempting to map a dependence network, this is helpful because it divides a complex inference into smaller, more manageable steps, showing the structure of the dependence. Consider prompting an LLM to analyse the dependence of a specific policy recommendation on underlying ethical principles in a bioethics paper using CoT: the output might explicitly state "Step 1: Identify the core ethical principle P (e.g., autonomy). Step 2: Identify the factual claim F about the situation. Step 3: Show how applying P to F leads to intermediate judgment J. Step 4: Show how J supports policy recommendation R." This makes the dependence of R on P, F, and J explicit and traceable. Observing such a chain allows the user to evaluate the coherence and validity of each step, leading to refinement of their own representation of the inferential structure. The potential for LLMs to integrate external tools also enhances their utility. Some systems can interact with calculators, databases, code interpreters, or formal logic provers (Millière & Buckner 2024). This allows for the verification of specific types of dependence claims. For instance, a philosopher working on modal logic could potentially use an LLM integrated with a prover to check the validity of a complex argument involving nested modal operators, verifying the claimed logical dependencies. Someone working in philosophy of science might use an LLM connected to scientific databases to quickly check empirical claims about causal dependencies relevant to a philosophical argument about scientific explanation. A decision theorist could use an LLM with a calculator or code interpreter to verify the dependence of an expected utility calculation on specific probability and utility assignments in a complex scenario. This integration can enhance the accuracy of specific dependence claims presented to the user, helping construct a more reliable representation of the dependence network, particularly where philosophy intersects with formal or empirical domains. It allows the LLM to function not just as a processor of existing text but as a way to use external validation sources for certain kinds of dependencies. In summary, these capabilities together – inferential competence derived from linguistic patterns, potentially rich latent representations and implicit world models capturing semantic and structural relations, effective context handling through self-attention, reasoning externalisation via techniques like CoT, plus the ability to leverage external tools – enable LLMs to generate text that can be input for a human user engaged in the process of refining their understanding by mapping dependence networks more accurately and comprehensively. These mechanisms operate primarily at the level of processing information structure as expressed in language and data, suited to the task of representing the often abstract dependencies central to philosophical inquiry. The effective application of these tools requires engagement from the philosopher, who must guide the process, evaluate the output, and integrate the insights gained into their own evolving understanding. The role of the user is central; the LLM functions as an assistant in the task of mapping complex relational structures. Continued research, particularly into interpretability and controlling LLM outputs, will likely further increase their utility in philosophical contexts. # Old Draft Dellsén et al. (2024) consider understanding to be a matter of the representation of dependence relations. This account differs from analyses of the ordinary meaning of 'understanding'; its stated aim is instead "to explicate it in a way that makes it suitable for being deployed in an account of philosophical progress" (Dellsén et al., 2024, p. 674). The central claim holds that the degree to which a subject understands a phenomenon corresponds to the accuracy and comprehensiveness of their representation of the network of dependence relations relevant to that phenomenon. These dependence relations are seen as objective, worldly relations that often ground explanatory practices. As Dellsén et al. describe, using Kim's (1994) work as a basis: > The idea that understanding is centrally concerned with representing dependence relations dates back to Kim (1994). For Kim, dependence relations are the ontological correlates of explanation – they are the worldly relations that make it true that something explains something else. In much of empirical science, the paradigmatic dependence relation is causation... It is a matter of contention which other relations are genuine dependence relations, but they may include constitution, grounding, mereological dependence, truthmaking, conceptual containment, and/or supervenience. (Dellsén et al., 2024, p. 675) The framework does not commit to an exact list of dependence relations, assuming that understanding involves representing whichever relations actually exist. Furthermore, understanding a phenomenon X requires grasping its position within a wider network representation; it extends beyond merely knowing the factors upon which X depends. > [...] understanding X is a matter of representing both how X depends on various other phenomena, and how further phenomena depend on X itself. In other words, the extent to which one understands X is a matter of how one represents the network of dependence relations running both to, and from, X. (Dellsén et al., 2024, p. 675) Such a representation includes facts about both existing dependencies (positive dependencies) and dependencies that do not exist (negative dependencies, for instance, that X lacks dependence on Y). Understanding, according to this view, admits of degrees. Its extent relates to two criteria: accuracy and comprehensiveness. Accuracy concerns how correctly the representation depicts the dependencies and non-dependencies that actually exist. Comprehensiveness concerns how fully the representation includes all relevant phenomena and the dependence relations (or their absence) connecting them. A tension may arise between these criteria. This could lead to trade-offs, for example idealisation (reducing accuracy for greater comprehensiveness) or abstraction (reducing comprehensiveness for greater accuracy) (Dellsén et al., 2024, p. 675). This account requires accuracy and comprehensiveness relative to the facts, which makes it "robustly factive". At the same time, it is described as "epistemically undemanding" (Dellsén et al., 2024, p. 674, 676): > [...] on the explication of understanding with which we will operate, understanding X does not imply having the type of epistemic justification that is required for knowing any particular proposition about X. [...] More importantly for our purposes, an explication of understanding that does not imply justification is arguably better suited for spelling out a plausible understanding-based account of philosophical progress. (Dellsén et al., 2024, p. 676) Thus, understanding in this specific sense does not require justification or belief regarding the represented dependencies. This notion—understanding as the accurate and comprehensive representation of networks of dependence relations, without requiring justification—guides the subsequent analysis of its development. Enhancing philosophical understanding, on the account presented previously, involves refining one's representation of the relevant network of dependence relations. When an individual encounters philosophical arguments or information that challenge their existing representation, cognitive processing occurs. This processing focuses on evaluating whether incorporating the implications of the new information would yield a network representation that is more accurate or more comprehensive. Greater accuracy means correctly mapping the dependencies that exist. Greater comprehensiveness means including a wider range of relevant dependencies and non-dependencies, potentially clarifying relations to and from key elements, or specifying positive/negative relations more fully. Imagine you want to enhance your understanding of a particular area of philosophy: _the nature of time_. You might read books or articles in philosophy journals concerning debates over A-series versus B-series theories, or temporal passage. This engagement prompts refinement of your dependence network representation. An initially inaccurate map linking passage solely to change might become more accurate by distinguishing subjective passage (A-series dependent) from objective ordering (B-series dependent). Comprehensiveness increases as the network incorporates dependencies involving metaphysics or physics, and dependencies _from_ passage to agency. Clarity improves by mapping negative dependencies (e.g., B-series ordering _not_ depending on subjective passage) or specifying dependence types (conceptual vs. metaphysical). Based on this processing, the individual adjusts their mental model of the dependence network, reflecting the gradable nature of understanding. The aim is to modify the representation so that it better maps the factive landscape of logical, conceptual, or explanatory relations. This might involve adding, removing, or altering specific dependencies. Consistent with the epistemically undemanding nature of this account, the resulting, revised representation constitutes enhanced understanding, even if the subject lacks justification or belief regarding every element within it. The enhancement lies in the improved accuracy or comprehensiveness of the representation itself. This refined representation is then integrated into the individual's broader philosophical framework. How much a given source enhances an individual's understanding depends on both their existing understanding and the quality of the source itself. For instance, engaging with a clear, well-informed article on A-series/B-series theories might significantly refine one's representation of time's dependence network, increasing its accuracy and comprehensiveness. Conversely, a poorly written or misleading article on the same topic might hinder this refinement, potentially leading to a less accurate representation. The degree to which understanding is enhanced—the improvement in the representation's accuracy or comprehensiveness—will therefore differ based on source quality, influencing the resulting map of dependencies, even if that representation remains epistemically undemanding. ## 2. LLMs, Understanding and Dependence Relations Recent developments in artificial intelligence, specifically the emergence of Large Language Models (LLMs), present possibilities regarding the enhancement of philosophical understanding. LLMs are computational systems built upon large neural networks, most commonly the Transformer architecture, trained on extensive datasets of text and code. Their primary function involves predicting subsequent linguistic tokens, enabling them to generate coherent and contextually relevant text in response to user prompts. Key to their operation is the self-attention mechanism, which allows the model to weigh the significance of different parts of the input context when producing output. Many of their complex behaviours are considered emergent properties arising from the scale of the models and the vastness of their training data. Several functional aspects of LLMs, derived from their architecture and training, are relevant when considering their potential to aid in representing dependence relations. They internalise and can reproduce the statistical patterns of language used to express dependencies (e.g., conditional statements, causal claims, explanatory structures) found in their training data. Their internal representations, often conceptualised as vector space embeddings, capture semantic relationships between concepts, which might reflect or suggest underlying conceptual or constitutive dependencies. Their context-handling abilities allow them to track information and relationships across segments of text. Furthermore, techniques like Chain-of-Thought (CoT) prompting can elicit step-by-step reasoning, potentially externalising inferential dependencies. LLMs can also synthesise information reflecting the broad world knowledge contained in their training data, potentially listing factors relevant to a phenomenon. Some systems can also integrate external tools, allowing for interaction with calculators, databases, or code interpreters to potentially verify specific types of dependence claims. An awareness of these features provides a basis for analysing how text generated by LLMs might assist a human user in representing dependence relations more accurately or comprehensively, according to the account of understanding outlined previously. ### How LLM Features Might Facilitate Representing Dependence Relations The specific capabilities of LLMs may contribute to a user's ability to build a more accurate and comprehensive map of dependence relations in several ways: - **Statistical Learning and Pattern Completion:** LLMs learn common linguistic structures that express dependence from their training data (e.g., "X causes Y", "P implies Q", "A is a necessary condition for B"). When prompted, they can generate text articulating standard philosophical arguments, definitions, or explanations by completing these learned patterns. This externalisation of known dependency claims (like the dependence of an action's moral status on its utility within utilitarianism) can help a user accurately represent established theoretical dependencies within a specific domain. It primarily aids the **accuracy** of representing known parts of the dependence **network**. - **Vector Space Semantics:** The way concepts are represented in an LLM's internal vector space, where related concepts might cluster together, can potentially suggest non-obvious dependencies to a user. Concepts situated closely might share constitutive elements or grounding relations. The LLM's capacity for analogical reasoning (identifying parallel structures like A:B :: C:?) might highlight similar dependency structures across different philosophical problems or even different disciplines. While any dependencies suggested this way require careful human evaluation for **accuracy** and **factivity**, this feature could enhance the **comprehensiveness** of the user's exploration by pointing towards potentially relevant factors or relationships within the **network** that were previously unconsidered. - **Self-Attention and Context Integration:** The self-attention mechanism enables LLMs to track relationships between terms and concepts across moderately long passages of text. This is relevant for representing dependencies accurately within complex arguments where, for example, a conclusion depends on premises stated several sentences or paragraphs earlier, or where the dependence involves specific entities introduced previously. By maintaining contextual coherence, the generated text is more likely to correctly represent the intended dependencies relative to the provided context, contributing to the **accuracy** of the user's map of that specific argumentative or explanatory **network**. - **Chain-of-Thought (CoT) and Emergent Circuits:** Prompting techniques like CoT encourage LLMs to break down reasoning into explicit steps. This process externalises the intermediate inferential dependencies, making the path from premises to conclusion more transparent. For a human user, this explicit articulation can enhance the **accuracy** of their understanding of that specific inferential path within the broader logical **dependence network**. Observing the generated chain allows the user to trace the claimed dependencies step-by-step. Identifying flaws or points of incoherence in the chain can also prompt critical refinement of the user's own representation of the dependencies involved. - **Knowledge Representation (Latent World Models):** LLMs synthesise vast amounts of information from their training data. When prompted about the factors influencing a phenomenon or the components of a concept, they can generate text drawing on this broad, latent knowledge base. This can help a user ensure a wider range of potentially relevant factors, and their associated dependencies, are considered, thereby aiding the **comprehensiveness** of their dependence **network** representation. The reliability of this depends on the **accuracy** of the LLM's synthesised knowledge. - **Tool Use Integration:** LLMs capable of interfacing with external, specialised tools can potentially provide more reliable information about specific dependencies. For example, interaction with a formal logic prover could verify logical entailment (a type of dependence). Accessing scientific databases could provide empirical evidence relevant to causal dependencies. Using calculation tools could verify quantitative dependencies. This integration can enhance the **accuracy** and **factive** basis of specific dependence claims presented to the user, supporting the construction of a more reliable representation of the dependence **network**. These features suggest mechanisms by which interaction with LLMs could, in principle, assist a human user in the process of refining their representation of dependence networks, aiming for greater accuracy and comprehensiveness as defined in the adopted account of understanding. All of the LLM features mentioned above—statistical learning, semantic representation, context integration, elicited reasoning, knowledge synthesis, and tool use—have either emerged or significantly improved in the last few years. Given the analysis in the preceding subsection, which links these specific features to the facilitation of representing dependence relations, such improvements should logically enhance the capacity of LLMs to generate text conducive to this process. Specifically, if LLMs become better at accurately reproducing dependency-expressing patterns, representing semantic relationships, tracking context, externalising reasoning steps, synthesising relevant knowledge, and verifying claims via tools, then their outputs should function as more effective instruments for users aiming to map dependence networks accurately and comprehensively. Consequently, newer generations of LLMs might, in principle, produce philosophical texts or engage in dialogues that are more effective aids for enhancing a user's understanding, defined as achieving more accurate and comprehensive representations of philosophical dependence networks. As an anecdotal observation, this potential for improvement seems apparent in practice. Engaging with LLMs on philosophical topics, for instance discussing ideas within an uploaded text in 2025, often yields noticeably different and more useful results compared to similar interactions in 2022 or 2023. While subjective, this common user experience lends intuitive support to the claim that the underlying technological capabilities relevant to philosophical exploration have indeed advanced. This potential depends on the continued development of the relevant LLM capabilities and assumes that the generated outputs are critically engaged with by the human user, as discussed previously regarding source quality and cognitive processing.