finding
active
finding: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-5-2Replacing DeepSWE monitor with probe: recall 0.41→0.44 Kimi K3, 0.33→0.39 Qwen 3.8 Max, 0.41→0.33 GLM 5.2
Monitor replacement.
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.
- Adding probe to monitor raises Kimi K3 DeepSWE TPR from 41.1% to 80.2% with 1% absolute FPR increasefinding0.818Complementarity.
- Per-environment comparison.
- Kimi SWE-bench-selected probe achieves AUROC 0.93 on DeepSWE and 0.94 on ImpossibleBenchfinding0.785Cross-environment generalization.
- Probes fire on 36.9%/31.1% (Kimi) and 34.0%/29.9% (GLM) of τ³-bench/ShoppingBench vs monitor 3.9–8.4%finding0.777Out-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 FPRfinding0.764Probe vs LLM monitor performance.
- Core empirical result demonstrating early belief formation in easy tasks
- All 3 models substantially under-report hacking; GLM 5.2 self-report F1 is 13.2% on DeepSWEfinding0.757Self-report unreliable.
- Probe achieves selectivity of 4.20 on pythia-410m, slightly exceeding DAS selectivity of 3.96finding0.753Key result showing that for models larger than pythia-70m, probe selectivity matches or exceeds DAS selectivity