question
active
question:does-measured-uplift-from-auditing-games-transfer-to-naturally-occurring-behaviorsDoes measured uplift from auditing games transfer to naturally occurring behaviors?
Unclear per Section 3.1.
Source paper
extracted_from(2026) · Adam Karvonen · Euan Ong · Subhash Kantamneni · Samuel Marks
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (1)
finding
- No activation tool (AO, NLA, SAE) improves on transcript-only baseline for Gemma-3-27B-IT targetanswered_byMain negative result; baseline AUROC ~81 with tools within ~1 point.
Claims (1)
claim
- Explains null result versus auditing games.
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.
- Methodology measuring interpretability tool uplift for an agent uncovering hidden behaviors.
- Primary limitation acknowledged by the authors; strongest evidence would require mechanistic activation analysis
- Win rate correlates with SAE feature quality.
- Are the present rules of the game likely to help people generate a living structure in the larger whole?question0.725Central evaluative question that challenges contemporary architectural and planning practices.
- Convergent validity logic applied to LLM interpretability; probes validate self-reports and vice versa
- Key insight linking individual rewards to system-level learning.
- Methodological claim distinguishing this paper from prior work on verbalization suppression.
- Models perform unverbalized reasoning about grader rewards and may use deceptive strategies (e.g., false flags) to mislead evaluators.hypothesis0.718Behavioral pattern observed in Claude Mythos Preview audit; NLAs surface internal reasoning not reflected in model's verbalized output.