claim
active
claim:the-choice-of-pelt-detection-cost-has-little-end-to-end-effectThe choice of PELT detection cost has little end-to-end effect.
Ablation.
Source paper
extracted_from(2026) · Eric Bigelow · Amir Zur · Satchel Grant · Tal Haklay +6
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (2)
finding
- Ablation.
- Bug discovered.
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.
- Used to segment base path at forking points with exact multinomial cost.
- With smoothed data, spacing and S trace the same Pareto frontier of accuracy vs token cost.finding0.715Cost tradeoff.
- The optimal layer for the prefill introspection differs from the optimal layer for detecting injected thoughts.
- Claude Opus 4.1 and 4 show greatest reduction in apology rate in the prefill detection taskfinding0.707Injecting a concept matching the prefilled word reduces the rate at which the model apologizes, maximally for Opus models.
- Choosing token indices to sample adaptively, as in optimal experiment design, could be more efficient than post-hoc smoothing.hypothesis0.702Future work.
- Feature presence depends on concept frequency in training data, with a threshold scaling inversely with alive features.
- Assumption D.3 formalized in the theoretical framework; empirically validated with coin-flip sequence experiments