A new arXiv preprint puts a closed form number on the delay after which AI 'referee' agents flip a multi agent LLM network from consensus to oscillation, and shows that grounding critics in retrieved facts suppresses the instability entirely.
Multi-agent AI systems with verifier or critic agents can flip from consensus to oscillation when a referee verdict arrives after the next round of argument. The paper, posted to arXiv on 25 June 2026, names a closed-form delay threshold for that flip and proves a fix: grounding critics in retrieved facts makes truth an absorbing boundary that eliminates the instability.
For builders of multi-agent LLM systems, which are pipelines where several language models debate, critique, or verify one another, the practical lesson is that more verification is not always better. There is a computable ceiling on how much correction and how much delay a system can absorb before it tips into oscillation, where agents stop converging on a shared answer and start contradicting each other round after round. Cross that ceiling, and the critics intended to keep the group honest become the engine of a permanent argument.
The model behind that ceiling is graph-based. The authors treat a multi-agent debate as a network of agents exchanging beliefs, with a subset of grounded corrector nodes standing in for verifiers or critics that can pull neighbors back toward fact. Delays are modeled directly: each link has a communication delay, and each grounded corrector introduces a verification delay before it returns a corrected signal. The paper's central tool is a spectral decomposition of the grounded Laplacian, a standard graph operator that here encodes how corrections propagate through the network, which yields a closed-form inequality separating stable consensus from oscillatory disagreement.
The headline number falls out of that analysis. For a two-step delay on both communication and verification, the closed-form stability threshold equals the inverse golden ratio, roughly 0.618. In plain terms, the worst oscillation regime is reached when the two delays line up, that is, when a critic's verdict arrives about as late as the next round of argument does. Below that dose-delay combination, false claims still stabilize, because grounded correctors can catch them in time. Above it, the system settles into a cycle instead of a conclusion.
The paper also gives designers a tool for staying under that ceiling when the corrector budget is finite. It frames the problem of where to place a limited number of grounded verifiers as a supermodular placement objective, a class of optimization problems where adding more candidates strictly improves the objective, and proves a greedy algorithm that achieves a (1 − 1/e) approximation, about 63 percent of the optimal placement. Translated out of the math, the rule is: spend your fact-checker budget on the most influential nodes, not on a few obvious ones, and you recover most of the value of a fully staffed verification layer.
The experiments test this prediction rather than just describe it. Across five open models, the authors confirm the dose-delay oscillation pattern predicted by the spectral analysis: when correction strength is too high or too late relative to communication delay, agents stop agreeing with one another and start cycling through contradictory answers. The match between theory and simulation is what gives the threshold its weight for engineers.
The most important result, however, is the counterpoint buried in the same paper. When verifiers are grounded in retrieved facts rather than in the agents' own prior beliefs, truth becomes an absorbing boundary, a state the system can enter but not leave, and the oscillation regime disappears entirely. The instability is not an unavoidable cost of multi-agent verification. It is a property of self-referential correction loops, and it goes away as soon as the critics have something external to point at.
That distinction matters for anyone deciding whether to add more critics or to retriever-ground the critics they already have. A second late-arriving verifier can deepen disagreement; a verifier anchored in a document store tends to end it. The paper argues that grounding is a structural fix for multi-agent instability, not a stylistic preference, and that the failure mode the threshold describes is specific to ungrounded correction.
The result is bounded by what the preprint actually proves. The model is theoretical, validated on five open-weight language models rather than on production systems, and the spectral analysis assumes the network can be represented as a grounded graph with clean delay parameters. The authors do not claim their threshold maps directly onto enterprise deployments, and the code release associated with the work has not been surfaced in this source pass. The headline number, 0.618, is exact only at delay two; other delay regimes have their own thresholds, all derivable but none as compact.
For builders, the watch items are narrow. If you run a multi-agent LLM pipeline with verifier or critic agents, measure the round-trip delay between an agent's claim and its critic's verdict, and check whether it lines up with the communication delay between agents. If the two delays are close, the threshold is at its tightest and the system is most likely to oscillate. Adding more critics in that regime is the move the paper specifically warns against; retriever-grounding the critics you have is the move it recommends.