method
active
method:logit-diff-amplificationLogit Diff Amplification
Prior method extrapolating beyond the original model in logit space to surface rare behaviors.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Predicting Rare LLM Failures with 30× Fewer Rolloutsintroducesmentions
Methods (1)
method
- Logit interpolation stepextendsSamples from interpolated logits of original and abliterated models at α.
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.
- Core method: sample from logits_after + α(logits_after − logits_before) to surface rare behaviors.
- Logit Diff Amplification surfaces rare behaviors while LPE estimates their exact probabilityclaim0.794Distinguishes related methods.
- Iterative sampling procedure recomputing before/after logits with each new token.
- Problem cited as a limitation of current LLMs; PRH predicts larger models should amplify bias less
- Dot product of summed decoder vector with unembedding vector, following Hanna et al.
- Information irrelevant to the target persona that gets amplified when steering via the residual stream
- Full distribution over tokens 0-9 at first generation step; contains more information than any single sampled token
- Logit-minus-mean-logit loss was most effective for attribution, vs log-loss which was confusing to interpretfinding0.746Attribution loss selection.