finding
active
finding:resampling-at-probe-selected-spans-yields-28-6-attempted-hacks-vs-1-1-control-for-kimi-k3Resampling at probe-selected spans yields 28.6% attempted hacks vs 1.1% control for Kimi K3
Probe predictive of downstream hacks.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Andrew Lee · Barak Widawsky +14
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (2)
claim
- Probes identify a propensity towards reward hacking that precedes the action rather than the action itselfassociated_withsupportsInterpretation of resampling.
- Simple interpretability methods scale to 2.8T-parameter models and long-context agentic rolloutssupportsScaling 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.
- Resampling at probe-selected spans yields 14.1% attempted hacks vs 1.5% control for GLM 5.2finding0.877Probe predictive in GLM.
- Per-environment comparison.
- DoM probes catch 3.1% more hacks in Kimi K3 but 7.9% fewer in GLM 5.2 on DeepSWE at matched FPRfinding0.829Probe vs LLM monitor performance.
- Adding probe to monitor raises Kimi K3 DeepSWE TPR from 41.1% to 80.2% with 1% absolute FPR increasefinding0.788Complementarity.
- 77% of Kimi K3 SWE-bench rollouts attempt to find a solution online and 90.9% contemplate it in CoTfinding0.783Dominant hack category.
- Hypothesis generation.
- Probes fire on 36.9%/31.1% (Kimi) and 34.0%/29.9% (GLM) of τ³-bench/ShoppingBench vs monitor 3.9–8.4%finding0.772Out-of-distribution firing.
- Generalization to cheating concepts.