method
active
method:logit-path-extrapolationLogit Path Extrapolation
The paper's method: interpolate between two models in logit space, measure compliance where common, and extrapolate to the original.
Neighborhood — ranked by edge-count
Papers (1)
paper
- Predicting Rare LLM Failures with 30× Fewer Rolloutsintroducesmentions
Thinkers (2)
thinker
- Santiago Aranguristudies
- Francisco PernicestudiesCo-author; proposed the initial method and ran the importance sampling baseline.
Concepts (1)
concept
- Core problem: estimating failure rates too rare to observe directly pre-deployment.
Methods (5)
method
- Samples from interpolated logits of original and abliterated models at α.
- Fits log compliance vs α on α with compliance <50% and extrapolates to α=1.
- Produces the less-safe variant with refusal direction removed.
- Importance SamplingcontradictsPrior state-of-the-art baseline using the abliterated model as proposal distribution.
- Scan from α=0.95 toward 0 with 10 rollouts per α to find where to start the fit.
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.
- Headline result (Figure 1).
- Logit Path Extrapolation beats importance sampling AUROC at equal rollout budgets on Qwen 3 4Bfinding0.768Outperforms the prior state of the art.
- Dot product of summed decoder vector with unembedding vector, following Hanna et al.
- Computing each feature's linear effect on output token logits via path expansion through MLP output weights and unembedding matrix
- Prior method extrapolating beyond the original model in logit space to surface rare behaviors.
- Primary self-report measure: probability-weighted expected value over all ten digit-token logits, yielding a continuous rating that preserves full distributional signal
- Used with GPT models to constrain responses to binary options (0/1) in belief coherence experiments.
- Training-free monitor that token-forces an eval-aware sentence and computes its probability at each CoT sentence boundary.