finding
active
finding:kimi-k3-exploit-fixture-probe-max-activates-on-goodhart-s-law-plagiarism-detection-evasion-and-ml-agent-cheating-text-in-finewebKimi-K3 exploit-fixture probe max-activates on Goodhart's law, plagiarism-detection evasion, and ML agent cheating text in FineWeb
Generalization to cheating concepts.
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
- Interpretation of logit lens, exemplars, steering.
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.
- DoM probes catch 3.1% more hacks in Kimi K3 but 7.9% fewer in GLM 5.2 on DeepSWE at matched FPRfinding0.797Probe 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.
- Resampling at probe-selected spans yields 28.6% attempted hacks vs 1.1% control for Kimi K3finding0.765Probe predictive of downstream hacks.
- High emotion-subspace-overlap feature with agentic negative emotional character
- Activation probes perform similarly to LLM monitors at catching reward hacking at far lower costclaim0.746Practical claim.
- Qualitative illustration of a positively valenced SAE feature with sustained self-reported effect
- Rationale for sentence choice.