NormAct: A Benchmark for Hidden Social Norm Compliance in Embodied Planning

A new benchmark called NormAct reveals that leading multimodal language models struggle to comply with unspoken social norms during embodied planning tasks, even when they successfully achieve explicit goals, according to research posted on arXiv [1]. The study, submitted June 26, introduces NormAct as a tool to evaluate how well AI planners handle hidden social constraints in egocentric environments [1]. Existing evaluations typically focus on explicit goal achievement or direct norm knowledge, seldom assessing whether planners can infer and apply these hidden constraints within action sequences [2]. NormAct embeds hidden norms within ordinary tasks, testing whether models can realize them without explicit instruction [1]. Experiments with three state-of-the-art multimodal large language models — GPT-5.4, Claude Opus 4.7, and Gemini 3 Pro — showed a pronounced gap. The models achieved explicit goals in 67.3% of cases, but complied with hidden norms in only 26.4% of cases [1]. Cue-condition experiments indicated that this gap stems not from a lack of general social knowledge, but from challenges in activating and grounding relevant norms in context [2]. To address the shortfall, the researchers propose NormPerceptor, a context-conditioned cue generator that infers scene-relevant norms prior to planning [1]. With NormPerceptor, overall Task Success rose from 24.2% to 46.7% [1]. The benchmark is publicly available on Hugging Face [2]. The paper appears on arXiv, an open-access repository of electronic preprints that are moderated but not peer reviewed [6]. Founded in 1991, arXiv now hosts over two million articles and receives about 24,000 submissions per month [6]. The platform also supports community-built tools through its arXivLabs framework, which allows collaborators to develop features such as citation explorers and code finders that appear on article record pages [4][5].