Explicit trust guidance improves AI negotiation agreement rates in preprint

Researchers report in a preprint that explicit guidance on using trust estimates increased agreement rates by up to 23 percentage points in repeated negotiations between language model agents. Agents that identified a disruptive participant did not consistently reach better agreements without that guidance. The researchers tested recurring agents negotiating over multiple issues while one participant covertly tried to prevent agreement. After each game, agents assessed their counterparts and carried trust estimates into later negotiations. The improvement depended on how agents assessed trust and incorporated it into their decisions.