A Close Look At World Model Recovery In Supervised Fine-Tuned LLM Planners

arXiv 2026 47.3 theory

TLDR

Investigates world model recovery in supervised fine-tuned LLM planners via interpretability experiments, finding linear encoding of action validity and state predicates.

Reasoning

Strengths: Novel interpretability approach to understand how LLMs represent planning knowledge. Weaknesses: Limited to classical planning problems; no real-world validation or benchmarks.

Read-first score

Read-first score 47.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 19.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Citation impact 18%

81.6

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.81613584

Methodology quality 18%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,experiment

Reproducibility 18%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Topical relevance 29%

27.1

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 293.

Keyword Scores

world model

world dynamics prediction

model-based reinforcement learning world model

world simulator

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Devised a series of interpretability experiments to interrogate world model recovery in fine-tuned LLM planners
Showed that supervised fine-tuning on valid action sequences enables linear encoding of action validity and some state predicates
Demonstrated that internal representations can separate valid from invalid actions even when output probabilities fail to classify them
Found that broader state space coverage (e.g., random walk data) during fine-tuning improves the accuracy of world model recovery
Contributed a recipe for applying interpretability techniques to planning LLMs

Methodology

The authors performed interpretability experiments on supervised fine-tuned LLMs for classical planning. They examined internal representations using linear probes and generative capabilities via output probabilities, comparing models fine-tuned on valid action sequences versus those with broader state space coverage (e.g., random walk data). The study evaluated action validity classification and state predicate encoding.

Key Results

Supervised fine-tuning on valid action sequences enables LLMs to linearly encode action validity and some state predicates. Models that fail to classify action validity using output probabilities may still have internal representations that separate valid from invalid actions, and broader state space coverage during fine-tuning yields more accurate recovery of the underlying world model.

Limitations

The study focuses on classical planning problems, which may limit generalizability to more complex or real-world planning tasks.
Interpretability analysis relies on linear probes, which may not capture non-linear or more complex representations.
The findings are based on specific fine-tuning data distributions (valid action sequences, random walks) and may not hold for other data types.

Resources

Paper