Researchers have found a smarter way to settle disagreements between AI agents than simply counting hands.
Most multi-agent LLM systems pick a final answer by majority vote: run the same problem through several agents, and whichever answer comes up most often wins. A new paper describes AgentAuditor, a method that instead builds a "Reasoning Tree" mapping where agents agree and where they split off from each other. When agents diverge, the system compares the actual evidence behind each branch rather than just counting how many agents landed on each answer. The team also trained a companion technique, called Anti-Consensus Preference Optimization, to stop the system from siding with the majority just because it's the majority. Tested across four multi-agent frameworks and several reasoning benchmarks, AgentAuditor beat both plain majority voting and LLM-as-judge setups, with accuracy gains up to 5 percentage points, while staying token-efficient.
The finding matters because majority voting has a real blind spot: agents built on similar training data can share the same biases and confidently agree on the same wrong answer, a failure mode the paper calls "confabulation consensus." As companies lean on multi-agent setups for coding, research, and decision-support tasks, an aggregation method blind to shared bias just means more confident wrong answers, faster.
Still, this is a single paper's benchmark results, not a deployed product, so treat the 5% figure as a lab number until someone reproduces it outside the authors' own test suite.