Free-Form Communication Is Why Language Models Outperform Bayesian Agents in Negotiation

In a solver-verified five-party negotiation scenario where language models and exactly specified Bayesian negotiators occupy interchangeable seats, tables of language models reach agreement far more often, in 0.958 of games for five Claude Opus 5 seats against 0.233 for five Bayesian negotiators. Agreement requires unanimity, so closing a deal is a coordination problem and the ability of LLMs to engage in free-form talk independently of proposing or accepting deals is what allows them to perform better. Three tests support this account: making the five Opus seats mute, so they still move but send no messages, costs 0.200 in agreement rate, as much as seating one mute Bayesian agent. The second test asks what the talk channel must carry; when each seat truthfully reveals its preferences once, every all-Bayesian table reaches agreement, whereas signals that stop short of revealing preferences plateau at an agreement rate of 0.400. The same declaration protocol adds less as language-model seats replace Bayesian ones, because free-form talk already conveys preferences. Attaching a language-model interface that speaks and listens for the Bayesian seat lets its table reach agreement as often as an all-Opus table, with the Bayesian policy still conditioning every proposal. A mute Bayesian seat lowers agreement at every capability tier, and an omniscient Bayesian seat helps Haiku tables but hurts Opus tables, so an algorithmic seat helps only when it knows something the table lacks and can communicate it. Designers should give every seat a free-form communication channel, wrap any mute policy such as an algorithmic participant in a language-model interface, and not seat a policy that knows nothing the table lacks, since the defaults otherwise cost a fifth of agreements. Experimental data and the interaction framework will be released on publication.