Towards Communication-Efficient Social Intelligence in Language Agents
TACT trains language agents to reach social goals with fewer words, using teacher revisions that test both strategy and expression.
The paper frames efficiency as more than shorter replies: an agent’s wording can change the partner’s next response and the rest of the exchange. Its method rewrites student actions with two specialists, one trimming unnecessary detail and one proposing better strategies for a partner’s constraints. Candidate revisions are judged by sampled partner responses and token cost, then distilled back into the student for standalone use. The authors report stronger goal results on SOTOPIA and better goal success with fewer tokens and messages on AgentSense. ArXiv · AI/CL/LG's note
The paper frames efficiency as more than shorter replies: an agent’s wording can change the partner’s next response and the rest of the exchange. Its method rewrites student actions with two specialists, one trimming unnecessary detail and one proposing better strategies for a partner’s constraints. Candidate revisions are judged by sampled partner responses and token cost, then distilled back into the student for standalone use. The authors report stronger goal results on SOTOPIA and better goal success with fewer tokens and messages on AgentSense. ArXiv · AI/CL/LG's note
score 4