Running queries through a panel of competing AI models and making them reach a consensus cuts hallucination rates by 41.7%.
Researchers have published a framework called Council Mode that sends a query to multiple frontier LLMs in parallel, then passes all their outputs to a dedicated consensus model. That third model maps where the panel agrees, where it diverges, and what unique findings each contributor surfaced. Tested on a 1,200-sample subset of the HaluEval benchmark under controlled, no-web conditions, the framework reduced hallucinations by 41.7% relative to the best individual model and gained 7.5 points on TruthfulQA. On a custom multi-domain reasoning benchmark, it scored 95.4% — a 9.2-point jump over the top solo model.
The result matters because it offers a path to higher factual reliability without retraining any model. Most hallucination-reduction work focuses on fine-tuning or retrieval augmentation; Council Mode treats the model zoo as a resource and uses disagreement between models as a signal, not a problem to hide. The approach also reportedly lowered measured bias variance, which is a harder claim to evaluate but worth watching.
The catch is cost: the framework burns 4.2x the token budget of a single-model call, which the authors freely admit makes it a poor fit for high-volume, low-stakes tasks. At that overhead, it slots into the same category as human-in-the-loop review — useful when being wrong is expensive, awkward elsewhere.