Benchmark RadarAI BENCHMARK PROFILE
NoRA
NoRA is a visual first-person video benchmark requiring models to generate candidate next actions and justify them via fact-reason-action support graphs. It includes 1,420 annotated video clips and uses a grounded reasonableness score combining action alignment, factual grounding, and support binding.
- Released
- 2026-06-03
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Normative competence in agents requires generating reasonable actions from scratch, grounded in visual facts. NoRA measures this ability and exposes gaps in current VLMs' reasoning.
Motivation
LLMs and agentic systems are increasingly deployed in social environments, making normative competence critical for safe and appropriate behavior.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.