finding
active
finding:kimi-swe-bench-selected-probe-achieves-auroc-0-93-on-deepswe-and-0-94-on-impossiblebenchKimi SWE-bench-selected probe achieves AUROC 0.93 on DeepSWE and 0.94 on ImpossibleBench
Cross-environment generalization.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Andrew Lee · Barak Widawsky +14
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Simple DoM vectors coherently represent reward hacking and are generalizable and interpretableassociated_withsupportsCentral claim.
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.
- Per-environment comparison.
- Adding probe to monitor raises Kimi K3 DeepSWE TPR from 41.1% to 80.2% with 1% absolute FPR increasefinding0.799Complementarity.
- Probe achieves selectivity of 4.20 on pythia-410m, slightly exceeding DAS selectivity of 3.96finding0.786Key result showing that for models larger than pythia-70m, probe selectivity matches or exceeds DAS selectivity
- Replacing DeepSWE monitor with probe: recall 0.41→0.44 Kimi K3, 0.33→0.39 Qwen 3.8 Max, 0.41→0.33 GLM 5.2finding0.785Monitor replacement.
- Moderate agreement.
- Probes fire on 36.9%/31.1% (Kimi) and 34.0%/29.9% (GLM) of τ³-bench/ShoppingBench vs monitor 3.9–8.4%finding0.769Out-of-distribution firing.
- DoM probes catch 3.1% more hacks in Kimi K3 but 7.9% fewer in GLM 5.2 on DeepSWE at matched FPRfinding0.762Probe vs LLM monitor performance.
- Likely-trained MM probe is a surprisingly effective causal baseline due to correlation between truth and probability on sp_en_trans