SOTOPIA-$\pi$: Interactive Learning of Socially Intelligent Language Agents

TL;DR

This paper introduces SOTOPIA-$ackslash pi$, an interactive learning approach that enhances social skills in language agents by combining imitation, social interaction, and self-reinforcement, leading to improved social and safety capabilities.

Contribution

The paper presents a novel interactive training method for language agents that improves social intelligence and safety, bridging a gap in social skill learning for AI.

Findings

A 7B LLM achieves social goal completion comparable to GPT-4-based agents.

The training improves safety and maintains general question-answering abilities.

LLM-based evaluations tend to overestimate social skills of trained agents.

Abstract

Humans learn social skills through both imitation and social interaction. This social learning process is largely understudied by existing research on building language agents. Motivated by this gap, we propose an interactive learning method, SOTOPIA-$\pi$, improving the social intelligence of language agents. This method leverages behavior cloning and self-reinforcement training on filtered social interaction data according to large language model (LLM) ratings. We show that our training method allows a 7B LLM to reach the social goal completion ability of an expert model (GPT-4-based agent), while improving the safety of language agents and maintaining general QA ability on the MMLU benchmark. We also find that this training paradigm uncovers some difficulties in LLM-based evaluation of social intelligence: LLM-based evaluators overestimate the abilities of the language agents trained specifically for social interaction.