claim
active
claim:icl-and-activation-steering-both-control-llm-behavior-by-updating-belief-in-latent-conceptsICL and activation steering both control LLM behavior by updating belief in latent concepts
Core thesis.
Source paper
extracted_from(2025) · Eric Bigelow · Daniel Wurgaft · YingQiao Wang · Noah D. Goodman +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (6)
finding
- Phase boundary jointly modulated by context length and steering magnitude, reconstructed by the Bayesian model with high fidelityassociated_withPrediction 3 confirmed.
- Non-persona generality.
- Cross-model generalization.
- Main predictive result.
- Generality test.
- Scale test.
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 likelihoodclaim0.828Mechanistic split.
- Nuanced interpretive claim about the limits of steering as a mechanism for reflection enhancement.
- Prediction 3.
- Scaling Laws for Activation Steering with Llama 2 Models and Refusal Mechanisms (Ali et al., 2025)concept0.773Related work finding larger models more resistant to steering, potentially consistent with ESR in 70B
- Steering derived from LRH.
- Central claim of the paper; supported by the model organism ground-truth approach.
- Limitation.
- Core mechanism statement.