Unverified50% confidenceFactExact time
A logprob decrease by itself cannot distinguish feature-specific suppression from magnitude artifacts—as long as the activation perturbation is large enough, it does not matter whether the direction is aligned with the target feature.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhen the intervention effect of the random-direction control is better than that of carefully selected SAE features, it indicates that the logprob decrease likely comes from perturbation magnitude rather than the semantic role of the feature itself.77% similarUnverifiedlogprob measures how fluent it is to generate that token, which is not equivalent to how correct the fact is; a model can perfectly well generate a false statement fluently and with high probability, so it remains a relative signal rather than an absolute guarantee.71% similarUnverifiedThere is no universal formula for choosing logit bias values; it is recommended to start with small adjustments, test the effect with representative prompts while watching for side effects, and then adjust gradually.68% similarUnverifiedAdvantage-sign-aware Renyi preprocessing is used to counteract saturation at the logit level, enabling gradient signals to propagate effectively in high-confidence regions and penalizing confident but incorrect predictions.68% similarUnverifiedModifying logits at sampling time means adding or subtracting bias values dimension by dimension on the vector before softmax—adding a positive value raises the probability of a certain type of token being selected, while subtracting a negative value suppresses it—without changing the model weights and intervening only at the inference stage.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133782API
curl https://kongchang.com/api/v1/knowledge/claims/1133782MCP
get_claim(id=1133782)