Towards Communication-Efficient Social Intelligence in Language Agents

TL;DR

This paper proposes Teacher-Assisted Communication Training (TACT) to improve social goal attainment in language agents while reducing communication cost. TACT revises student-generated actions and expressions to preserve intended actions while removing unnecessary details. It achieves state-of-the-art results on SOTOPIA and AgentSense benchmarks.

Why It Matters

Effective communication in language agents is crucial for real-world applications, such as customer service and negotiation. TACT's approach can lead to more productive and less demanding interactions between humans and language agents. This can improve user experience and increase the adoption of language agents in various industries.

Key Findings

• TACT improves social goal attainment while reducing communication cost
• TACT outperforms existing methods on SOTOPIA and AgentSense benchmarks
• TACT's expression specialist and strategy specialist contribute to its effectiveness