finding
active
finding:linear-probes-at-layer-9-predict-behavioral-beliefs-with-rmse-09-emotions-09-genres-11-arbitrary-on-llama-3-1-8bLinear probes at layer 9 predict behavioral beliefs with RMSE .09 (Emotions), .09 (Genres), .11 (Arbitrary) on Llama-3.1-8B
Beliefs linearly decodable.
Source paper
extracted_from(2026) · Eric Bigelow · Raphaël Sarfati · Daniel Wurgaft · Owen Lewis +4
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (3)
claim
- LLMs assign beliefs over a low-dimensional geometric conceptual belief space, and ICL corresponds to a trajectory through itassociated_withCentral thesis.
- Finding (2).
- Validates behavioral probe.
Findings (1)
finding
- Layer selection.
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.
- Likely-trained MM probe is a surprisingly effective causal baseline due to correlation between truth and probability on sp_en_trans
- Intrinsic noise in representation.
- Linear readout works but is not the true representation.
- At Llama layer 26 (~81% depth) the predictive component does not dominate the latent norm.finding0.796Fig. 14.
- App. E.
- Core E3 finding validating S as a predictor of anchoring effectiveness
- Third promising case from temporal permutation analysis.
- Empirical demonstration that MDVP produces divergent representations in a real LLM