Knowing Whom to Trust is not Enough: The Trust--Agreement Gap in Repeated Multi-Party LLM Negotiation

[arXiv] Amit Sarker, Moumita Choudhury, Saaduddin Mahmud and others propose a repeated multi-party, multi-issue LLM negotiation framework: multiple repeatedly participating agents engage in a series of negotiation games, in which one participant secretly obstructs reaching an agreement; after each round, agents evaluate their opponents and retain pairwise trust estimates. After comparing trust-aware and trust-free agents, the study finds a so-called "trust–agreement gap": trust-aware agents can identify the saboteur within a few rounds, but accurate trust estimates do not always translate into better agreements; only when agents are explicitly guided to use trust in their decisions does the agreement rate increase by up to 23 percentage points. The authors state that explicit trust can improve repeated LLM negotiations, but its value depends on how trust is evaluated and incorporated into agent reasoning.