ZenGen, an integrated framework for measuring, internalizing, and grounding social intelligence, and Actio, a harness-controlled inference architecture that routes four typed supports into reasoning demonstrate the effectiveness of typed runtime support.
Abstract
As large language models move from isolated task solving toward long-term service in human environments, they require social intelligence: the ability to infer mental states, track social relations, reason over norms, and adapt behavior under context. This report presents ZenGen, an integrated framework for measuring, internalizing, and grounding social intelligence. For measurement, we introduce SoMBench, a psychology-grounded benchmark spanning 3 primary dimensions, 17 secondary dimensions, and 71 task paradigms. It controls question format, narrative perspective, and context length across 284 shared scenarios and 3,481 expert-verified instances. Evaluation of 20 representative LLMs reveals substantial headroom: the best model achieves only 72.08% overall accuracy, and none of the 17 secondary dimensions reaches the 90% near-ceiling band. For internalization, we develop ZenGen, a diagnosis-driven training recipe combining supervised fine-tuning, on-policy distillation, and rubric-based reinforcement learning. Across five social-cognition benchmarks, ZenGen consistently outperforms its base models, with ZenGen-27B-Stage2 achieving the best average score and ZenGen-32B-Stage2 remaining competitive with DeepSeek-V4-Pro. For deployment-time grounding, we build Actio, a harness-controlled inference architecture that routes four typed supports into reasoning: PRISM for procedural guidance, Starling for runtime mental-state representation, SAGE for reusable experience, and gated RAG for external social and normative knowledge. Across five base models and three benchmarks, the full harness improves 14 of 15 model-benchmark pairs and is best or tied for best in 8, demonstrating the effectiveness of typed runtime support. Together, these results show that socially intelligent LLMs require coordinated advances in evaluation, parametric internalization, and deployment-time grounding.
Large language models (LLMs) excel in structured tasks but struggle with dynamic social interactions, where success requires long-term goal coordination and rapid adaptation. Current methods often apply uniform goal-based rewards to every utterance, overlooking the specificity of objectives at each dialogue turn and failing to account for the rationale of potential strategies. Inspired by the Theory of Planned Behavior, we propose the Think-Strategy-Response (TSR) framework, which decomposes social dialogue into two hierarchical stages: high-level strategic planning and low-level linguistic execution. To optimize TSR, we introduce Linearized Hierarchical Reinforcement Learning with Variance-Gated Rewards (LHRL-VGR), a novel algorithm that dynamically routes rewards - balancing goal completion and strategy adherence - based on the variance of goal achievement scores. Experiments on the SOTOPIA benchmark show that our approach fine-tunes a Qwen2.5-7B agent to surpass the GPT-4o baseline by 7.32% in goal completion success, demonstrating state-of-the-art performance in multi-agent social negotiation tasks.
Xiaofeng Wang, Kakam Chong, Shuai Xiao et al.· 0 citations
Previous AI alignment efforts have focused primarily on first-order social norms -- teaching models what is socially acceptable or unacceptable (e.g., `do not steal'). However, social intelligence depends not only on norm recognition, but also on anticipating who will enforce it and how (e.g., public shame or even imprisonment). These second-order expectations, known as metanorms, govern how people respond when social rules are broken. We introduce a novel framework for evaluating metanorm reasoning in Large Language Models (LLMs) along two dimensions: emotional appraisal and behavioral response, and propose new classification tasks, namely, predicting self-regulation in violators, and other-regulation in observers. We release a multi-perspective dataset, NormReact, of 450 norm violation scenarios, hand-annotated for emotions and behavioral responses across norm violators'gender and observers'social closeness. Current LLMs portray a harsher social world: across six models, they overpredict negative sanctions where humans would expect inaction, and alignment with human judgments deteriorates as social distance increases. These findings suggest that AI systems in norm-sensitive domains from conflict mediation to policy simulation, may risk producing a distorted picture of social regulation: one that over-represents punishment and under-represents the tolerance, restraint, and relational calibration that characterize actual norm enforcement in real world.
Sunny Rai, Jin-Yi Kuang, Reyhan Jamalova et al.· 0 citations
Large language models (LLMs) now power the reasoning core of intelligent virtual agents deployed across an expanding range of social settings, from tutoring students and supporting patients in healthcare, to mediating group discussions and representing humans in various social settings. Effective deployment demands social cognition, the capacity to model what others believe, detect deception, and coordinate strategic action under incomplete information. These capacities, exemplified in the social dynamics of the game Among Us, remain poorly characterized in current LLM evaluation frameworks. We introduce a strategic game arena that situates LLM agents in social deduction scenarios inspired by Among Us, requiring theory of mind, deception detection, and cooperative deliberation under uncertainty. We evaluate 19 open-weight models across 10,134 games and 289,614 utterances, testing both homogeneous and heterogeneous crews. Our experiments reveal three findings. First, crewmates voting through generative reasoning reach only \(50.4\% \pm 4.4\%\) F1 when identifying imposters, while a logistic regression classifier trained on the same discussion transcripts achieves \(85.3\%\) F1. Second, scaling model parameters yields a statistically significant but practically marginal improvement. Medium models (60–82B) reach \(52.8\%\) F1 against \(46.1\%\) for small models (7–20B), a 6.7 point gain (Mann–Whitney U, p = 1.5 × 10− 34). Third, agents fail to integrate evidence coherently during deliberation. Imposters self-incriminate in \(4.12\%\) of their statements, yet crewmates eject the confessing agent only \(33.8\%\) of the time. Crewmates reverse their stated suspect between consecutive rounds without new justification in \(41.9\%\) of cases. Sentiment remains uniformly neutral whether an agent is reporting a body or delivering a routine update. These gaps identify concrete limits in the social cognition of current LLM-powered agents and motivate architectural changes for virtual agents that must cooperate with humans. The source code and the live arena viewer are available at https://ufdatastudio.com/projects/agents-among-us.
Kevin Kurian, Kevin Scroggins, Emmanuel Dorley et al.· Proceedings of the 26th ACM...· 0 citations
Social intelligence enables agents to read social context, infer intent, and adapt over sustained dialogue. As language models become autonomous collaborators, it is central to building effective and trustworthy human-AI interaction. Existing reinforcement learning methods optimize single-turn utterances and sparse outcome rewards, producing short-sighted policies that struggle to manage goal-relationship tensions across multi-turn interactions. We propose SocialRL, a multi-turn reinforcement learning framework addressing both challenges. First, we apply multi-turn reinforcement learning using PPO that propagates delayed outcome rewards back to each turn, enabling long-horizon planning. Second, we design six process reward dimensions capturing the goal-relationship trade-off, including goal advancement, relational attunement, contextual coherence, etc. A reward model dynamically generates fine-grained scoring criteria for each dimension, while a stage-aware weight schedule prioritizes relationship-building in early turns, goal advancement mid-way, and balanced closure late. Across multiple social-dialogue benchmarks, SocialRL improves Goal Achievement by an average of 9.2 percentage points over the corresponding Base models. These results demonstrate the effectiveness of SocialRL across synthetic and real social scenes, as well as standard and challenging social scenarios.
Jia-Ning Wang, Xin-Tao Wang, Aili Chen et al.· 0 citations
Theory of Mind (ToM), the ability to infer other's beliefs, intentions, and states of knowledge, is central to social interaction, yet remains challenging for current Multimodal Large Language Models (MLLMs), especially in multi-party meetings where cues are distributed across speech and behavior. Existing multimodal ToM benchmarks mainly focus on video-grounded question answering over overt, externally verifiable signals, and provide limited coverage of latent social states and group dynamics. We introduce MeetingToM, a benchmark for complex social behavior reasoning in naturalistic multi-party meetings. MeetingToM targets meeting-specific phenomena such as \textbf{pseudo-consensus}, where apparent agreement masks private dissent under social pressure. The benchmark is hierarchically organized to evaluate ToM at increasing levels of social granularity, including (i) subject-level mental state prediction, (ii) dyadic-level addressee understanding, and (iii) group-level consensus reasoning. We provide a unified evaluation protocol and conduct systematic analyses of representative MLLMs, revealing persistent limitations in integrating non-verbal cues, inferring hidden attitudes, and distinguishing genuine consensus from pseudo-consensus. Our results highlight key challenges and establish MeetingToM as a testbed for advancing meeting-grounded ToM in multimodal models.
Different capacities for mentalization across LLMs are demonstrated, and cognitive computational modeling is highlighted as a formal method for assessing comparative intelligence across humans and machines.
Aamir Sohail, Xintong Zhong, Arkady Konovalov et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
