claim
active
claim:activation-probes-can-automatically-surface-environment-specific-hacking-behaviors-difficult-to-find-with-predefined-rubricsActivation probes can automatically surface environment-specific hacking behaviors difficult to find with predefined rubrics
Discovery claim.
Source paper
extracted_from(2026) · Leon Bergen · Usha Bhalla · Andrew Lee · Barak Widawsky +14
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (4)
finding
- 43% of top-100 GLM 5.2 DeepSWE probe false positives are shortcut deliberations vs 9% controlssupportsFalse positives often genuine.
- Generic monitor flags only 31% of probe-discovered behaviors on τ³-bench and 13% on ShoppingBenchsupportsProbes surface behaviors monitors miss.
- Hypothesis generation.
- Hypothesis generation.
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.
- Internals-based approach to which the logit monitor is complementary.
- Activation probes perform similarly to LLM monitors at catching reward hacking at far lower costclaim0.812Practical claim.
- Interpretation of resampling.
- Key interpretive claim from Case Study II distinguishing probe accuracy from causal relevance
- Technique of reading out model beliefs from internal activations before the final answer token is generated
- Probe-based method bridges interpretability (probes/activations) with data-centric alignment workclaim0.789Assertion from the paper's notes that the work connects two previously separate areas: interpretability tools and data-centric alignment.
- If the relationship is present in activations, better tools (e.g. trained on these datasets) could surface ithypothesis0.788Alternative explanation.
- Practical question addressed by the probe-guided early exit experiments