Anthrodial creates more human-like social interaction for language models
AnthroDial: Benchmarking LLM Anthropomorphism in Autonomous Social Interaction
Computation and Language
Summary
It's hard for AI agents to interact socially like humans because they must know when and how to communicate naturally over time. The authors created AnthroDial, a set of tools and methods that help language models chat more like people by managing conversations independently and adapting to changes. They also made ways to measure how well these AI agents understand thoughts, feelings, and behaviors during talks. Their tests showed AnthroDial helps models act more naturally and makes improvements easy to spot. This gives a solid foundation for building better AI conversational partners.
What this means in practice
- •For chatbot developers: Create AI chatbots that decide when and how to talk naturally in ongoing conversations using AnthroDial's tools and training methods.
- •For game designers: Design game characters that interact autonomously and adapt to players across long and varied dialogues based on AnthroDial’s interaction and evaluation techniques.
Authors
Wentao Liu, Xi Chen, Siyu Song, Biao Yuan, Yu Zhang, Zhou Zhuotong, Jingying Zhou, Guohao Feng, Shasha Hu, Tianfu Wang, Shangshang Yang, Haoyang Liu, Youjia Li, Xiaokun Wang, Min Ji, Ji Wang
Abstract
Large language models (LLMs) are increasingly deployed as social agents, yet credible human-like interaction requires more than fluent responses or persona consistency. Agents must autonomously decide whether, when, and how to communicate while adapting to evolving contexts, goals, and relationships. Existing research, however, lacks a unified approach to enabling, evaluating, and improving such capabilities in continuous, open-ended interaction. We introduce AnthroDial, a unified framework for developing anthropomorphic social agents from three complementary aspects: MindFlow, a lightweight interaction harness that enables autonomous, asynchronous, and adaptive communication through a dynamic Mind Buffer; CAPS-Eval, a theory-grounded framework for evaluating cognitive, affective, and behavioral dimensions of anthropomorphic interaction; and a scalable training paradigm that combines SEEDS for environment expansion with DiAPO for adaptive capability optimization. We further construct evaluation datasets covering everyday communication, game interaction, and long-horizon character interaction. Extensive experiments across diverse models and scenarios demonstrate improved interaction autonomy and naturalness, validate the reliability, discriminativeness, and agreement with human rankings of CAPS-Eval, and confirm the effectiveness of our training paradigm. Together, these components provide a unified framework for developing credible human-like social agents in open-ended interaction.