claim
active
claim:llms-assign-beliefs-over-a-low-dimensional-geometric-conceptual-belief-space-and-icl-corresponds-to-a-trajectory-through-itLLMs assign beliefs over a low-dimensional geometric conceptual belief space, and ICL corresponds to a trajectory through it
Central thesis.
Source paper
extracted_from(2026) · Eric Bigelow · Raphaël Sarfati · Daniel Wurgaft · Owen Lewis +4
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Linear probes at layer 9 predict behavioral beliefs with RMSE .09 (Emotions), .09 (Genres), .11 (Arbitrary) on Llama-3.1-8Bassociated_withBeliefs linearly decodable.
Claims (6)
claim
- Finding (3).
- Finding (2).
- Finding (1).
- Theoretical aim.
- Overall conclusion.
- Motivation.
Questions (1)
question
- Future work.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- ICL and activation steering both control LLM behavior by updating belief in latent conceptsclaim0.797Core thesis.
- Conclusion from Experiment 2 on Leap-of-Thought.
- The authors' interpretive assertion based on their steering results.
- Primary positive claim of the paper, grounded in strength comparison and localization results
- Core claim directly challenged by prior work denying introspection; forms foundation for Koan Battery introspection studies.
- Interpretation of the layer-by-layer PCA visualizations showing linear structure emerging in early-middle layers
- Theoretical hypothesis about the mechanism underlying LLM error detection and reflection.
- Central thesis statement of the paper