finding
active
finding:base-monitor-achieves-30-9-0-9-overall-reductionBase + Monitor achieves 30.9%±0.9% overall reduction
Comparable to RLFR-NI.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Model features can serve as inexpensive, dense, scalable supervision for open-ended tasks.associated_withCentral thesis.
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.
- Binary detection adjusted accuracy reaches 97.3% at layer 0 with α=5 before baseline control is appliedfinding0.770The misleadingly high result that prior paradigm would report as evidence of introspection
- Base corrections: 72.9% Failed Fix, 14.7% Fixed; retractions: 69.4% Correct Retract, 30.0% Incorrect Retractfinding0.761Label distributions.
- SFT on correction+retraction data gives 39.4% inline overall reduction and 9.6% not-inline, well below RLFRfinding0.750SFT baseline comparison.
- Ablation of BoN.
- Dropping 50% of monitor calls on SWE-bench with GLM lowers precision 0.656→0.588 while Kimi drops ~0.01finding0.746Cost of replacement.
- Quantifies how much of the base model's diversity VS can recover compared to baseline prompting
- Intervention action distribution.
- Ablation removing inline interventions.