Unverified50% confidenceFactExact time
Refusal behavior is largely mediated by a single direction in the residual stream, typically computed as the mean difference vector between harmful and harmless prompts.
1
Sources
50%
Confidence
Long-term
Relevance
10/22/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen fine-tuned on samples of safe refusals, the model learns refusal as a tendency, which may lead it to over-refuse legitimate questions as well.79% similarUnverifiedThe criterion for judging whether results can be skimmed rather than examined closely is tiering by impact scope and reversibility: low-risk, easily rolled-back results can be approved, while results involving core logic or irreversible operations must be manually confirmed.73% similarUnverifiedThe evidence gap is defined as the probability of accepting a correct correction minus the probability of yielding to pressure toward a wrong answer, normalized to a 0–1 interval, drawing on the psychological framework distinguishing informational influence from normative influence.71% similarUnverifiedA refusal-to-answer mechanism triggers a degradation strategy—such as returning an uncertain response, requesting human review, or narrowing the scope of the answer—when it detects that the input is out of the training distribution or the model's confidence is below a threshold.71% similarUnverifiedThe study proposes using a causal counterfactual criterion to suppress false starts: if muting the user input leaves the model's output distribution nearly unchanged, the event is identified as a false start and suppressed.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1084575API
curl https://kongchang.com/api/v1/knowledge/claims/1084575MCP
get_claim(id=1084575)