Cross-Model Consensus Outperforms Reward Models for LLM Reasoning

Selecting the most accurate answer from a pool of candidate reasoning chains is critical for scaling Large Language Model (LLM) performance at test-time. Current selection methods, such as self-consistency or trained reward models, each have drawbacks: self-consistency inherits the errors of a single model, while reward models require labeled data and often struggle with out-of-distribution generalization. This paper explores a novel, inference-time free signal: cross-model consensus. The "LLM-jury" approach treats a panel of independently trained models as a jury, where the structure of their agreement, rather than any individual model's score, serves as the verification signal. Across seven benchmarks, this method consistently selects correct answers more effectively than self-consistency and significantly better than a model attempting to score its own candidates. For instance, on competition math problems, it nearly closes the entire gap to an oracle selector. The underlying mechanism is "error decorrelation": independently trained models tend to make different mistakes, causing incorrect answers to scatter while the correct answer accumulates agreement. A parameter-free law derived in the paper accurately predicts consensus accuracy and identifies a "shared-error floor" where models might collectively hold a misconception. Compared to four trained verifiers, the LLM-jury matches the strongest within their training domain and emerges as the superior selector outside it, offering a transparent and predictable verification method.