claim
active
claim:steering-operates-by-changing-concept-priors-while-icl-accumulates-evidence-via-likelihoodSteering operates by changing concept priors, while ICL accumulates evidence via likelihood
Mechanistic split.
Source paper
extracted_from(2025) · Eric Bigelow · Daniel Wurgaft · YingQiao Wang · Noah D. Goodman +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- Positive steering magnitudes shift ICL curves leftward and negative rightward on Llama-3.1-8BsupportsSteering alters ICL dynamics.
Claims (1)
claim
- Derived from LRH.
Quotes (1)
quote
- Core mechanism statement.
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.
- Prediction 3.
- ICL and activation steering both control LLM behavior by updating belief in latent conceptsclaim0.828Core thesis.
- Boundary between regimes dominated by c or c'.
- Addresses skeptical alternative that reports reflect only conversational content
- Mechanism claim supported by transcript analysis and the fact that the steering vector was extracted from a model that never writes type hints.
- Key advantage of preventative over post-hoc steering: lower side-effect cost on general capabilities
- Suggests localization.
- Validates using chain-of-thought belief monitoring as proxy for behavioral steering efficacy.