claim
active
claim:icl-and-steering-additively-affect-log-posterior-odds-yielding-distinct-phasesICL and steering additively affect log posterior odds, yielding distinct phases
Prediction 3.
Source paper
extracted_from(2025) · Eric Bigelow · Daniel Wurgaft · YingQiao Wang · Noah D. Goodman +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Prediction 3 confirmed.
- Cross-model phase boundary prediction.
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.
- Steering operates by changing concept priors, while ICL accumulates evidence via likelihoodclaim0.858Mechanistic split.
- Steering changes behavior regardless of input x, as a constant shift in log posterior oddsclaim0.838Derived from LRH.
- Prediction 3: log posterior odds additive in context and steering, with crossover N* predictable from prior odds and magnitudehypothesis0.823Eq. 10.
- ICL and activation steering both control LLM behavior by updating belief in latent conceptsclaim0.796Core thesis.
- Boundary between regimes dominated by c or c'.
- Positive steering magnitudes shift ICL curves leftward and negative rightward on Llama-3.1-8Bfinding0.767Steering alters ICL dynamics.
- Validates using chain-of-thought belief monitoring as proxy for behavioral steering efficacy.
- Mechanism claim supported by transcript analysis and the fact that the steering vector was extracted from a model that never writes type hints.