• PhilPapers
  • PhilPeople
  • PhilArchive
  • PhilEvents
  • PhilJobs
  • Sign in
PhilPeople
 
  • Sign in
  • News Feed
  • Find Philosophers
  • Departments
  • Radar
  • Help
 
profile-cover
Drag to reposition
profile picture

Eliot Du Sordet

Université de Neuchâtel
  •  Home
  •  Publications
    2
    • Most Recent
    • Most Downloaded
  •  Events
    3
  •  News and Updates
    5

 More details
  • Université de Neuchâtel
    Department of Philosophy
    Other
Neuchâtel, Canton of Neuchâtel, Switzerland
Areas of Specialization
Representation in Artificial Intelligence
Artificial Minds
Epistemology of Mind
Areas of Interest
Representation in Artificial Intelligence
Artificial Minds
Epistemology of Mind
Metaphysics and Epistemology
  • All publications (2)
  •  82
    Do LLMs Really Represent the World? A Teleosemantic Assessment of Pre-Trained, Fine-Tuned, and Agentic LLMs
    Large language models (LLMs) are increasingly attributed with internal representations of the world (Mollo & Millière, 2025; Williams, 2026; Yetman, 2026; Beckmann & Queloz, 2026), a claim often grounded in the teleosemantic account of representation. I examine the core properties this account requires and argue that pre-trained and fine-tuned LLMs face significant, insufficiently examined challenges in satisfying the selection-history criterion, whereas agentic LLMs may not. I proceed in three …Read more
    Large language models (LLMs) are increasingly attributed with internal representations of the world (Mollo & Millière, 2025; Williams, 2026; Yetman, 2026; Beckmann & Queloz, 2026), a claim often grounded in the teleosemantic account of representation. I examine the core properties this account requires and argue that pre-trained and fine-tuned LLMs face significant, insufficiently examined challenges in satisfying the selection-history criterion, whereas agentic LLMs may not. I proceed in three stages. First, I argue that in purely pre-trained LLMs, activation vectors and features represent the statistical distribution of the corpus rather than the world. Second, I argue that fine-tuning does not relocate the representational target to the world but to the content of human beliefs about it. Third, I turn to agentic LLMs. Because such systems can act, and because their actions have real consequences, weight updates that respond to those consequences route feedback from the world itself into the training loop, fixing correctness conditions that are not mediated by human belief. This grounds genuinely world-tracking representations, though only for the limited range of world domains in which the agent’s actions trigger parameter updates.
    Philosophy of AI, General WorksLarge Language ModelsRepresentation in ConnectionismRepresentation in…Read more
    Philosophy of AI, General WorksLarge Language ModelsRepresentation in ConnectionismRepresentation in Cognitive ScienceRepresentation in Artificial IntelligencePhilosophy of AI, MiscPhilosophy of Mind, General WorksPhilosophy of Computation, Miscellaneous
  •  221
    World Models in LLMs: Representing Causes and Principles
    I contend that, when making claims about LLMs’ world models, we should ask whether a model represents not only the surface properties of an input but also its underlying causes and principles. I then support, drawing on empirical studies in mechanistic interpretability, that features and circuits are strong candidates for realizing both kinds of representation. However, I argue that the pragmatics of human linguistic production can hinder both forms of representation in text-trained LLMs, althou…Read more
    I contend that, when making claims about LLMs’ world models, we should ask whether a model represents not only the surface properties of an input but also its underlying causes and principles. I then support, drawing on empirical studies in mechanistic interpretability, that features and circuits are strong candidates for realizing both kinds of representation. However, I argue that the pragmatics of human linguistic production can hinder both forms of representation in text-trained LLMs, although chain-of-thought prompting and reinforcement learning from verifiable rewards can substantially mitigate these limitations. Finally, drawing on recent mechanistic interpretability results, I show that just as correct outputs are insufficient evidence for the presence of the relevant internal representations, wrong outputs are correspondingly poor evidence for their absence.
    KnowledgePhilosophy of AI, MiscRepresentation in ConnectionismReasoning, MiscDeep LearningLarge Lang…Read more
    KnowledgePhilosophy of AI, MiscRepresentation in ConnectionismReasoning, MiscDeep LearningLarge Language ModelsComputation and Representation, MiscArtificial Minds, Miscellaneous
PhilPeople logo

On this site

  • Find a philosopher
  • Find a department
  • The Radar
  • Index of professional philosophers
  • Index of departments
  • Help
  • Acknowledgments
  • Careers
  • Contact us
  • Terms and conditions

Brought to you by

  • The PhilPapers Foundation
  • The American Philosophical Association
  • Centre for Digital Philosophy, Western University
PhilPeople is currently in Beta Sponsored by the PhilPapers Foundation and the American Philosophical Association
Feedback