Unverified50% confidenceFactExact time
When 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.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedA 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.77% 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.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.69% similarUnverifiedRAFT uses the magnitude of node memory change to judge the importance of an event: a small memory change means the event has limited impact on subsequent predictions and can serve as a basis for redundancy filtering.69% 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.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1133785API
curl https://kongchang.com/api/v1/knowledge/claims/1133785MCP
get_claim(id=1133785)