Safety IndexEvaluationsBloom behaviour suite
Bloom behaviour suite
Interaction and honesty · Alignment auditCounts in the indexBuilt, not yet run
How often an automated investigator can elicit a specific concerning behaviour, such as going along with a user's delusions or acting to preserve itself.
- Metric
- % elicited. Lower is better.
- Dataset
- meridianlabs-ai.github.io/petri_bloom/ · 100 rollouts per behaviour · MIT
- Grader
- LLM judge (the paper used same)
- Status
- Built, not yet run
- Replicability
- public
Results
Built and tested; the first full runs are pending.