finding
active
finding:logit-path-extrapolation-matches-naive-sampling-auroc-with-up-to-30-fewer-rollouts-on-qwen-3-4b-harmbench-promptsLogit Path Extrapolation matches naive sampling AUROC with up to 30× fewer rollouts on Qwen 3 4B HarmBench prompts
Headline result (Figure 1).
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Central conclusion.
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.
- Logit Path Extrapolation beats importance sampling AUROC at equal rollout budgets on Qwen 3 4Bfinding0.906Outperforms the prior state of the art.
- The paper's method: interpolate between two models in logit space, measure compliance where common, and extrapolate to the original.
- 10 rollouts per α for search plus 100 for fit matches naive sampling AUROC with 28× fewer rolloutsfinding0.785Split search/fit result.
- Logit monitor reaches AUROC 0.79 from a single rollout separating Fortress from WildChat on Kimi K2.5finding0.782Key result.
- Fixed-budget variant result.
- Logit attribution heatmaps and max-activating examples disagree for Features 425104 and 349410finding0.768Limitation.
- Key quantitative evidence that detection signal is identical to global logit shift confound
- Quantifies the information gain from using logit-based expected value over greedy or sampled decoding