ArXiv 2610.31971 measures 23 percentage point agreement lift only when LLM agents receive explicit trust instructions

The framework runs repeated multi-issue negotiations among fixed agents with one hidden saboteur. After each round agents update pairwise trust scores that persist across games. Trust-agnostic baselines and trust-aware variants are compared under identical payoff structures and communication protocols. Results isolate the trust-agreement gap: identification succeeds rapidly while outcome improvement requires an additional reasoning step that maps trust values to proposal and acceptance thresholds.

Classical multi-agent literature supplies explicit reputation mechanisms that propagate distrust into strategy updates; LLM work has treated trust as emergent from context alone. The paper shows emergence produces correct mental models without corresponding policy change. When the model is prompted to condition offers on stored trust values, agreement rises 23 points; absent that prompt the gap remains. This pattern matches observed failures in deployed LLM mediators that correctly flag unreliable counterparties yet still accept suboptimal terms.

Operational implication is immediate for any system fielding LLM negotiators on behalf of organizations. Logging trust alone adds negligible value; the decision layer must contain an explicit rule or fine-tuning objective that penalizes cooperation with low-trust agents. Without that rule, repeated interactions will continue to produce accurate diagnostics paired with unchanged loss rates.

Next deployments will test whether fine-tuning on trust-conditioned trajectories closes the gap more reliably than prompting. Controlled A/B trials measuring both adversary detection latency and final utility delta are required within six months of any production rollout.