Talk is Cheap, Communication is Hard: Dynamic Grounding Failures and Repair in Multi-Agent Negotiation

A new paper from researchers Yiheng Yao, Chelsea Zou, and Robert D. Hawkins (arXiv:2605.01750) challenges the assumption that large language models (LLMs) can effectively negotiate in multi-agent settings. The team designed an iterated negotiation game where two agents allocate shared resources toward private projects, with verifiable jointly optimal outcomes. While individual agents could identify Pareto-optimal allocations in isolation, agent dyads consistently failed to reach them across both open-source (e.g., Llama) and closed-source (e.g., GPT-4, Claude) models. The investigation reveals four distinct failure modes: (1) coordination degrades when agents lack shared interaction history; (2) accumulated context becomes a liability via stubborn anchoring, where initial proposals are treated as axiomatic; (3) agents default to perfunctory fairness (equal splits) over reward-maximizing coordination; and (4) referential binding fails—agents lose track of commitments across turns.

The study decomposes the coordination gap into measurable components. An oracle baseline shows the gap is not due to individual reasoning limitations; a no-talk baseline confirms communication is necessary; and a full-transparency intervention proves information exchange alone is insufficient. The bottleneck lies in the interactive processes of joint plan formation, commitment, and execution—what the authors call dynamic grounding. This contrasts with static grounding tasks common in current benchmarks. The findings highlight dynamic grounding as a critical, understudied axis for multi-agent coordination, with implications for deploying LLMs in real-world negotiations, collaborative planning, and autonomous systems where mutually beneficial outcomes depend on iterative communication and trust.