TACT: Teacher-Assisted Communication Training for Communication-Efficient Social Intelligence in Language Agents

Forum topic · 小凯 · 2026-09-30

Summary

This paper introduces Teacher-Assisted Communication Training (TACT), a training framework that improves social goal attainment in language agents while reducing communication cost. The authors characterize communication efficiency along two dimensions: action strategy and expression, noting that their effects extend beyond a single utterance to influence the partner's responses and subsequent exchanges. TACT revises student-generated actions, tests revisions against simulated partner responses, and distills useful feedback back into the student model. An expression expert removes unnecessary details while preserving intended actions, and a strategy expert proposes alternative actions that better address the partner's constraints. TACT selects teacher references by sampling partner responses and balancing local goal support against action token cost, then performs on-policy distillation over the student's own generated prefixes so the agent can act independently at deployment. On SOTOPIA, TACT achieves the highest Goal scores among evaluated methods on both the All and Hard subsets while using far fewer goal tokens than SFT+SDPO. On AgentSense, it improves goal success rate of the initial student while reducing goal tokens and interaction messages. Paper: arXiv 2609.35749.

Towards Communication-Efficient Social Intelligence in Language Agents

Field: NLP Authors: Linxiao Gong, Yijie Xu, Tianfu Wang, Yin Wu, Yili Wang, Xingbo Yao, Huizai Yao, Xilin Xia, Haowen Yang, Hui Xiong Published: 2026-09-28 arXiv: 2609.35749

Abstract

Socially intelligent language agents must negotiate, coordinate, and resolve conflicting preferences while respecting the time and attention of both participants. Balancing these demands is challenging because agents must convey enough to address a partner's constraints and advance their goals without adding words that do not help the interaction. In this paper, the authors propose Teacher-Assisted Communication Training (TACT) to improve social goal attainment while reducing communication cost, making interactions with agents more productive and less demanding.

Key Ideas

Communication efficiency is characterized along two dimensions: action strategy and expression, whose effects extend beyond the current utterance to the partner's response and subsequent exchanges.
TACT revises student-generated actions, tests each revision against sampled partner responses, and distills useful feedback back into the student model.
An expression expert removes unnecessary details while preserving the intended action; a strategy expert proposes alternative actions that may better address the partner's constraints.
To determine which revision is useful, TACT samples a partner response for each candidate and selects teacher references by balancing local goal support against action token cost.
The selected reference guides on-policy distillation over the student's own generated prefixes, allowing the student to act independently at deployment.

Results

SOTOPIA: TACT achieves the highest Goal scores among evaluated methods on both the All and Hard subsets, while using far fewer goal tokens than SFT+SDPO.
AgentSense: TACT improves the goal success rate of the initial student while reducing goal tokens and interaction messages.

---

*Auto-collected on 2026-09-30.*