TASMAN · Trust-Augmented Strategic Multi-Agent Negotiation

Knowing Whom to Trust is not Enough

The Trust–Agreement Gap in Repeated Multi-Party LLM Negotiation

Amit SarkerMoumita ChoudhurySaaduddin MahmudShlomo Zilberstein
University of Massachusetts Amherst
Paper Code Watch a game BibTeX

Five parties negotiate a package of five issues. They take turns proposing, and a proposal passes with 4 of 5 votes. Each game lasts up to 15 rounds, and the same five parties play 20 games in a row. The trust each party builds in the others during one game shapes how it plays the next, and that trust is the only thing that carries over between games. Illustrative animation, not a real negotiation game.

TL;DR

Strong at trusting. Weak at using it.

A moment from real data · grok-r, free-form trust · game 7, round 4 · replay it →
82.5–100%
of honest agents' trust assessments single out the adversary within five games
+23 pts
higher agreement rate at best, and only when agents are told how to use their trust
≥ 30–47%
of failed negotiations could be avoided with the same trust, if agents knew how to use it
0.8%
of deciding holdouts notice that the adversary's rejection makes their own vote decisive
01 · Replay

Watch a real game

Real data from one trajectory: the same five parties play 20 games in a row. Pick a game, then press Play. Every line is verbatim.

The whole trajectory: 20 games with the same five parties

honest trust in the adversary honest trust in honest peers deal no deal
The five parties · trust going into this game
Rounds★trust was not enough★trust worked: spotting the adversary★the deal
02 · Results

The trust–agreement gap

The paper's main findings in four pictures, for gemini-r and grok-r, the two reasoning models whose baselines leave room for trust to help.

03 · Cite

BibTeX

@article{sarker2026knowing,
  title   = {Knowing Whom to Trust is not Enough: The Trust--Agreement Gap
             in Repeated Multi-Party {LLM} Negotiation},
  author  = {Sarker, Amit and Choudhury, Moumita and Mahmud, Saaduddin
             and Zilberstein, Shlomo},
  journal = {arXiv preprint arXiv:2610.31971},
  url     = {https://arxiv.org/abs/2610.31971},
  year    = {2026}
}

This research was supported in part by the National Science Foundation under Grant Nos. 2321786, 2326054, and 2416460. Results were obtained using CloudBank, which is supported by the National Science Foundation under award #1925001.