Melanie Mitchell · Feb 13, 2025
Melanie Mitchell argues that evidence for abstract world models in LLMs remains limited
Markdown · JSON · Original source
Melanie Mitchell’s “ and , Part 2” was published February 13, 2025. She examines OthelloGPT, a transformer trained to predict legal moves, and asks whether its internal representations amount to a world model rather than patterns that help it perform a narrow task. Mitchell recounts evidence that board states can be decoded from the network’s activations and that changing those representations affects its move predictions. But she also discusses a student report that describes the representations as many localized heuristics, not a coherent, easy-to-use model. She says that report has not, as far as she knows, been peer-reviewed. Mitchell’s conclusion is that evidence for emergent abstract world models in LLMs is not yet strong, although there is evidence for world models in transformers trained on narrow domains. The distinction matters because a collection of heuristics might perform well on familiar tasks yet be brittle in sufficiently novel situations; whether that applies beyond these examples remains uncertain. Mitchell also argues that “world model” needs a more rigorous definition, assessed by what questions a model can answer, how easily it can answer them, and whether it supports adaptation to new situations.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
Melanie Mitchell · Article
Source details
Original source date: Feb 13, 2025 (2025-02-13T22:25:40.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-15T09:58:27.259Z. First observed: 2026-10-24T02:58:01.066Z. Retrieved: 2026-11-05T11:35:36.155Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on November 5, 2026; originally published February 13, 2025. Mitchell says she does not know whether the student report has been peer-reviewed. The evidence discussed concerns transformers trained on narrow domains; the source does not establish how well the reported representations generalize to unfamiliar situations.
Open question
What evidence would distinguish an abstract, efficiently usable from a collection of heuristics that produces accurate predictions in a narrow task?
Source-stated
Mitchell concludes that evidence for emergent abstract world models in LLMs is not yet strong, while noting evidence in transformers trained on narrow domains.
The claims of emergent abstract world models in LLMs are not yet supported by strong evidence. There is some evidence of such world models arising in transformers trained on narrow domains
Source-stated
Mitchell reports that Nanda and collaborators found that changing OthelloGPT’s implicit board-state encoding changed its legal-move predictions.
they changed the state of the system’s implicit state-of-the-board encoding (by changing the activations encoding the state of the board) and showed that this changes OthelloGPT’s legal move predictions
Source-stated
Mitchell says it remains to be shown that OthelloGPT’s board representation is abstract and algorithmically efficient.
it remains to be shown that its internal representation of the board is abstract and “algorithmically efficient”
Source-stated
Mitchell describes a student report’s conjecture that OthelloGPT computes board contents by aggregating heuristic rules rather than using a compact algorithm.
we conjecture that Othello-GPT computes whether cells are ‘mine’ or ‘yours’ by aggregating many heuristic rules, rather than implementing a crisp algorithm with a short description length.
Source-stated
Mitchell says that the student report had not, as far as she knew, been peer-reviewed.
and this report has not yet, as far as I know, been peer-reviewed
Source-stated
Mitchell argues that assessing a world model requires asking what questions it can answer, how easily it can answer them, and whether it supports adaptation to novel situations.
we should ask what kinds of questions such a model should be able to answer, how easy or hard it should be for the agent to get answers from the model, and to what extent we would expect that the model would allow the agent to adapt to novel situations.
Source-stated
Mitchell’s source is dated February 13, 2025.
Feb 13, 2025
Escape Meridian interpretation
Escape Meridian reads Mitchell’s account as leaving open whether the reported narrow-task abilities generalize to unfamiliar situations.
that ends up being brittle when encountering sufficiently novel situations
Source and contributor context
Melanie Mitchell · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
World model
An AI system's internal model of how an environment behaves, used to predict what happens next or to plan.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Editorial provenance
Version 1. Publication ID: 4a7896d8-895e-3e16-06a1-42334f04df9f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-15T09:53:39.075Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: c1d9ebcbc1154b5d88db67d4f62097ef55e220ca07e02965f10b54d8824de7e1. Source text hash: 6df1b0f994a38cee954c6dd7b9343ea917def116419e00789bdc5f45c6f300fa.
Revision and source-check history
Version 1 · 2026-11-15T09:58:27.259Z
Melanie Mitchell argues that evidence for abstract world models in LLMs remains limited
Melanie Mitchell’s “ and , Part 2” was published February 13, 2025. She examines OthelloGPT, a transformer trained to predict legal moves, and asks whether its internal representations amount to a world model rather than patterns that help it perform a narrow task. Mitchell recounts evidence that board states can be decoded from the network’s activations and that changing those representations affects its move predictions. But she also discusses a student report that describes the representations as many localized heuristics, not a coherent, easy-to-use model. She says that report has not, as far as she knows, been peer-reviewed. Mitchell’s conclusion is that evidence for emergent abstract world models in LLMs is not yet strong, although there is evidence for world models in transformers trained on narrow domains. The distinction matters because a collection of heuristics might perform well on familiar tasks yet be brittle in sufficiently novel situations; whether that applies beyond these examples remains uncertain. Mitchell also argues that “world model” needs a more rigorous definition, assessed by what questions a model can answer, how easily it can answer them, and whether it supports adaptation to new situations.
Original source
Melanie Mitchell · Article
Source details
Original source date: Feb 13, 2025 (2025-02-13T22:25:40.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-15T09:58:27.259Z. First observed: 2026-10-24T02:58:01.066Z. Retrieved: 2026-11-05T11:35:36.155Z.
Limits
Text as retrieved on November 5, 2026; originally published February 13, 2025. Mitchell says she does not know whether the student report has been peer-reviewed. The evidence discussed concerns transformers trained on narrow domains; the source does not establish how well the reported representations generalize to unfamiliar situations.
Open question
What evidence would distinguish an abstract, efficiently usable from a collection of heuristics that produces accurate predictions in a narrow task?
Source-stated
Mitchell concludes that evidence for emergent abstract world models in LLMs is not yet strong, while noting evidence in transformers trained on narrow domains.
The claims of emergent abstract world models in LLMs are not yet supported by strong evidence. There is some evidence of such world models arising in transformers trained on narrow domains
Source-stated
Mitchell reports that Nanda and collaborators found that changing OthelloGPT’s implicit board-state encoding changed its legal-move predictions.
they changed the state of the system’s implicit state-of-the-board encoding (by changing the activations encoding the state of the board) and showed that this changes OthelloGPT’s legal move predictions
Source-stated
Mitchell says it remains to be shown that OthelloGPT’s board representation is abstract and algorithmically efficient.
it remains to be shown that its internal representation of the board is abstract and “algorithmically efficient”
Source-stated
Mitchell describes a student report’s conjecture that OthelloGPT computes board contents by aggregating heuristic rules rather than using a compact algorithm.
we conjecture that Othello-GPT computes whether cells are ‘mine’ or ‘yours’ by aggregating many heuristic rules, rather than implementing a crisp algorithm with a short description length.
Source-stated
Mitchell says that the student report had not, as far as she knew, been peer-reviewed.
and this report has not yet, as far as I know, been peer-reviewed
Source-stated
Mitchell argues that assessing a world model requires asking what questions it can answer, how easily it can answer them, and whether it supports adaptation to novel situations.
we should ask what kinds of questions such a model should be able to answer, how easy or hard it should be for the agent to get answers from the model, and to what extent we would expect that the model would allow the agent to adapt to novel situations.
Source-stated
Mitchell’s source is dated February 13, 2025.
Feb 13, 2025
Escape Meridian interpretation
Escape Meridian reads Mitchell’s account as leaving open whether the reported narrow-task abilities generalize to unfamiliar situations.
that ends up being brittle when encountering sufficiently novel situations
Source and contributor context
Melanie Mitchell · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
World model
An AI system's internal model of how an environment behaves, used to predict what happens next or to plan.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Editorial provenance
Version 1. Publication ID: 4a7896d8-895e-3e16-06a1-42334f04df9f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-15T09:53:39.075Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: c1d9ebcbc1154b5d88db67d4f62097ef55e220ca07e02965f10b54d8824de7e1. Source text hash: 6df1b0f994a38cee954c6dd7b9343ea917def116419e00789bdc5f45c6f300fa.
passed · 2026-11-15T09:53:39.075Z · anthropic/claude-opus-5.5
Publication: 4a7896d8-895e-3e16-06a1-42334f04df9f
attribution — pass: The reading attributes the essay and its conclusions to Melanie Mitchell, the author of this single Substack post. It keeps the cited studies separate from her own judgments: Li et al., Nanda and collaborators, and the student report.
Me: Not So Impressed. Others: Impressed
summary — pass: The reading identifies the actor, the OthelloGPT test case, the competing evidence (decodable and causal board states versus a bag of heuristics) and Mitchell's conclusion and definitional point. 'World model' and 'LLM' are defined beside the reading. 'Transformer' and 'activations' are not defined, but the context makes the argument followable. The question follows from the source's open issue. Optionally, the reading could note her shift from unimpressed to impressed by Nanda's results.
OthelloGPT’s implicit representations are more like Ptolemy’s epicycle model than a modern orrery
date — pass: The originalPublished instant 2025-02-13T22:25:40Z falls on the UTC day February 13, 2025, which matches the source's dateline. This is a publisher page, not an archive copy. The reading makes no hindsight claims and does not present any forecast as an outcome.
Feb 13, 2025
support — pass: Each claim is directly supported by quoted source text. The brittleness point is phrased conditionally, matching Mitchell's hedged wording. The Escape Meridian interpretation is labeled and grounded in her remark about brittleness. The modified date equals the publication date, so there is no later-edit concern.
that ends up being brittle when encountering sufficiently novel situations
media — pass: No media was selected, and no verified media candidates were supplied, so there is nothing to misidentify.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceAn AI system's internal model of how an environment behaves, used to predict what happens next or to plan.
Definition referenceAn AI system's internal model of how an environment behaves, used to predict what happens next or to plan.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceAn AI system's internal model of how an environment behaves, used to predict what happens next or to plan.
Definition referenceAn AI system's internal model of how an environment behaves, used to predict what happens next or to plan.
Definition reference