NoRA: Evaluating Grounded Reasonableness in Visual First-person Normative Action Reasoning

Researchers introduce NoRA, a visual reasoning benchmark that evaluates whether AI models can generate and justify appropriate actions in first-person video scenarios through explicit reasoning graphs. The benchmark reveals that current multimodal language models struggle to construct complete action spaces and properly ground decisions in visible evidence, highlighting a critical gap between selecting plausible actions and explaining them through verifiable reasoning.

NoRA addresses a fundamental limitation in how AI safety and competence are currently measured. Existing normative reasoning evaluations either operate purely in text or force models to select from predetermined options—neither scenario reflects real-world deployment where agents must independently identify and justify actions. This benchmark shifts evaluation methodology by requiring models to generate candidate actions, support them with scene facts, and bind decisions to specific visual evidence through explicit reasoning graphs.

The research emerges from growing concerns about AI deployment in social environments where normative judgment directly impacts safety and appropriateness. As large language models and multimodal systems enter decision-making roles, understanding whether they can reason about appropriate behavior grounded in observable reality becomes increasingly critical. NoRA's 1,420 annotated video clips provide empirical measurement of this capability gap.

The benchmark's findings are particularly revealing: while 12 tested multimodal systems frequently identify plausible actions and relevant scene facts, they consistently fail at two crucial tasks—constructing comprehensive action spaces and binding selected actions to correct visual evidence. This disconnect between partial competence and full reasoning capability has significant implications for developers building agentic systems, suggesting current approaches may generate superficially reasonable outputs without authentic grounding.

For the AI industry, NoRA establishes measurable standards for evaluating normative competence in multimodal systems. The research direction suggests future model development must focus not just on action selection accuracy but on transparent, verifiable reasoning chains connecting visible facts to decisions. This methodology could influence how safety-critical AI systems are validated before deployment in interactive environments.

• →NoRA benchmark requires AI models to generate and justify actions from first-person video, not select from preset options, better reflecting real-world deployment scenarios.
• →Current multimodal language models identify plausible actions and relevant facts but fail to construct complete action spaces or properly ground decisions in visual evidence.
• →The benchmark introduces grounded reasonableness scoring combining action alignment, factual grounding, and support binding into a single comprehensive metric.
• →Testing across 12 VLMs reveals systematic weaknesses in binding selected actions to correct local visual support, indicating a fundamental reasoning gap.
• →NoRA shifts AI evaluation from binary action correctness to whether models can justify appropriate decisions through inspectable, verifiable reasoning chains.