finding
active
finding:kimi-k3-probes-catch-3-0-and-33-1-points-more-hacks-on-deepswe-and-impossiblebench-but-miss-6-6-more-on-swe-bench-at-matched-fprKimi K3 probes catch 3.0 and 33.1 points more hacks on DeepSWE and ImpossibleBench but miss 6.6 more on SWE-bench at matched FPR
Per-environment comparison.
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
- Activation probes perform similarly to LLM monitors at catching reward hacking at far lower costsupportsPractical 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.
- DoM probes catch 3.1% more hacks in Kimi K3 but 7.9% fewer in GLM 5.2 on DeepSWE at matched FPRfinding0.888Probe vs LLM monitor performance.
- Adding probe to monitor raises Kimi K3 DeepSWE TPR from 41.1% to 80.2% with 1% absolute FPR increasefinding0.843Complementarity.
- Resampling at probe-selected spans yields 28.6% attempted hacks vs 1.1% control for Kimi K3finding0.841Probe predictive of downstream hacks.
- Kimi SWE-bench-selected probe achieves AUROC 0.93 on DeepSWE and 0.94 on ImpossibleBenchfinding0.838Cross-environment generalization.
- Generalization to cheating concepts.
- 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 5.2finding0.792Monitor replacement.
- Suggests training on evaluation.
- Probes fire on 36.9%/31.1% (Kimi) and 34.0%/29.9% (GLM) of τ³-bench/ShoppingBench vs monitor 3.9–8.4%finding0.780Out-of-distribution firing.