Researchers have published four studies examining how large language models perform in multi-agent scenarios across diverse applications.
According to arxiv.org, a paper titled “Moral Hazard in Multi-Agent Language Models” introduces the Dialogue Moral Hazard Game, a controlled experimental environment testing whether AI agents cooperate when effort is costly and primarily benefits others. The study evaluated eleven open-weight language models and three frontier API models, finding sharp behavioral differences. According to the paper, “Fable 5 moves from querying toward local reward as cost rises,” while “GPT-5.6 Sol reaches ceiling behavior in the primary setting.” In incentive-isolation experiments with 3,015 decisions, Sol tracked a theoretical boundary “with a mean absolute error of 0.013.”
A separate arxiv.org paper, StreamReason-Bench, tests whether language models understand event-time stream processing. According to the research, models perform poorly on event-time reasoning: “no model that actually follows the instruction clears 34% exact match” when answering directly, though chain-of-thought reasoning “roughly doubles that for several of them.”
Two additional papers explore practical applications. According to arxiv.org, Foam-Agent automates computational fluid dynamics workflows in OpenFOAM using a multi-agent framework, achieving “an 88.2% execution success rate on the 110 Basic-tier tasks.” Meanwhile, arxiv.org reports that MAS-DecStream applies LLM-assisted contract negotiation to edge-cloud scheduling, “reducing latency violations to 3%” and improving utility by up to 22% over baselines.