method
active
method:resampling-at-probe-peak-spansResampling at probe-peak spans
Truncate and resample CoT at probe peaks versus matched controls to test prediction.
Neighborhood — ranked by edge-count
Papers (1)
paper
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.
- Resampling at probe-selected spans yields 14.1% attempted hacks vs 1.5% control for GLM 5.2finding0.803Probe predictive in GLM.
- Resampling at probe-selected spans yields 28.6% attempted hacks vs 1.1% control for Kimi K3finding0.750Probe predictive of downstream hacks.
- The paper's new probe: continuously varying a model's initial latent state along 2D slices and labeling outcomes by convergence time.
- Causal transformer probe predicting whether a token is in the same Entity as its predecessor.
- Probe quality on held-out set.
- Probe firing escalates to LLM monitor to lower FPR.
- Probe cosine similarity becomes more peaked with deeper layers 0-3, with higher secondary ringsfinding0.726Sharpening.
- Probe predictions are systematically miscalibrated until isotonic regression corrects biasfinding0.724Calibration necessity.