framework
active
framework:auditing-gamesAuditing games
Methodology measuring interpretability tool uplift for an agent uncovering hidden behaviors.
Neighborhood — ranked by edge-count
Papers (2)
paper
- Building and evaluating alignment auditing agentsintroducesmentions
Methods (2)
method
- Evaluation of tool uplift over transcript-only baseline using accuracy and AUROC.
- Investigation seeded with held-out meta-rhyme behavior.
Concepts (1)
concept
- Reward model sycophancyassociated_withHidden objective of the target model: exploiting believed RM biases.
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.
- Win rate correlates with SAE feature quality.
- Clear exemplar of process art; designed to engender aesthetic experiences of player choice, movement, and cognition.
- Unclear per Section 3.1.
- Interpretability tools and semantic search are key to winning the Marks et al. auditing gameclaim0.739Affordance insight.
- Systematic investigation of a model's alignment-relevant propensities.
- Aggregating 10 reports.
- Illustrative analogy showing that a dialogue agent never commits to a single object but maintains possible objects in superposition