claim
active
claim:steering-affects-type-hint-writing-by-modifying-the-model-s-belief-about-whether-it-is-being-evaluated-not-by-directly-encoding-type-hint-informationSteering affects type hint writing by modifying the model's belief about whether it is being evaluated, not by directly encoding type hint information
Mechanism claim supported by transcript analysis and the fact that the steering vector was extracted from a model that never writes type hints.
Source paper
extracted_from(2025) · Hua, Tim Tian · Qin, Andrew · Marks, Samuel · Nanda, Neel
Neighborhood — ranked by edge-count
Findings (3)
finding
- Steering with same vector on pre-fine-tuned Llama Nemotron has zero effect on type hint rateassociated_withsupportsFalsifies hypothesis that steering vector encodes information about type hints; steering can only affect behavior via deployment-belief modification.
- Validates using chain-of-thought belief monitoring as proxy for behavioral steering efficacy.
- Shows model does not use token-level matching to trigger type hints; correctly identifies that it must be evaluated, not a third party.
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 changes behavior regardless of input x, as a constant shift in log posterior oddsclaim0.808Derived from LRH.
- Core mechanism statement.
- Central interpretive claim organizing the entire paper's results
- Steering operates by changing concept priors, while ICL accumulates evidence via likelihoodclaim0.794Mechanistic split.
- The paper's critique of the standard linear steering baseline, supported by the days-of-week demo.
- Main result: steering elicits deployment behavior even when the evaluation cue is present and prompting fails.
- Limitation.
- Do latents for precise distinctions in types of behaviour have distinct steering effects?question0.787Authors have not verified this.