Benchmark RadarRevengeBench
RevengeBench is a benchmark for recovering code-space policies from behavioral traces. It includes 75 LLM-generated policies across five game environments, where a learner designs behavioral probes and submits executable hypotheses, evaluated using continuous action-distance metrics.
- Released
- 2026-06-24
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Addresses the inverse problem of inferring hidden decision programs from observations, relevant to opponent modeling and policy interpretability. It provides a tractable testbed for studying how controlled experiments improve code-space recovery.
Motivation
For most of scientific history, researchers studying behavior could only infer hidden mechanisms from outward actions: an inverse problem that becomes more tractable when observation is augmented by targeted intervention.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.