LLMs as a Jury: Cross-Model Consensus Can Outperform Process Reward Models for LLM Reasoning

2026-07-11 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

A new verification signal, "cross-model consensus," proposes using an "LLM-jury" to select correct answers from candidate reasoning chains, offering a free alternative at inference time. This method, which relies on the degree of agreement among independently trained models, significantly outperforms self-consistency and self-scoring models across seven benchmarks. Notably, on competition math, it closes the entire gap to an oracle selector, a feat self-scoring models barely achieve. The underlying mechanism is error decorrelation, where diverse errors scatter, allowing correct answers to accumulate agreement. A parameter-free law predicts consensus accuracy with a mean absolute error of 0.03, also identifying a "shared-error floor" that is minimal in math but present in science. This LLM-jury approach matches the strongest trained verifiers within their math training domain and surpasses them outside it.

Key takeaway

For Machine Learning Engineers developing LLM reasoning systems, consider implementing cross-model consensus as a robust, inference-time free verification signal. This approach, leveraging diverse LLMs, can significantly improve answer selection accuracy, especially in domains like competition math, often surpassing traditional self-consistency or trained reward models. You can characterize its reliability using the provided parameter-free law, understanding its shared-error floor to optimize its application.

Key insights

Cross-model consensus, leveraging error decorrelation among diverse LLMs, effectively verifies reasoning chains, often outperforming trained reward models.

Principles

• Independently trained models err differently.
• Error decorrelation enhances collective accuracy.
• Consensus accuracy is predictable via a parameter-free law.

Method

Independently trained LLMs each solve a problem once; their final answer agreement forms a cross-model consensus signal, verifying reasoning chains without additional training or labeled data.

In practice

• Verify LLM reasoning chains at inference time.
• Improve answer selection in competition math.
• Characterize verifier reliability using consensus law.

Topics

• LLM Reasoning
• Cross-Model Consensus
• Error Decorrelation
• Reward Models
• Inference Verification
• Competition Math

Best for: Research Scientist, AI Engineer, AI Architect, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Cross-Model Disagreement as a Label-Free Correctness Signal — Cross-model disagreement detects confident LLM errors by measuring a second model's surprise at the first's answer.
• To Mix or To Merge: Toward Multi-Domain Reinforcement Learning for Large Language Models — Multi-task RLVR for LLMs offers comparable performance to model merging with significantly reduced computational cost.
• What's the theoretical basis for using llm consensus as a probability estimator for real world events [R] — LLM consensus for event probability estimation faces challenges with correlated errors and out-of-distribution events.
• Reliable Chain-of-Thought via Prefix Consistency — Truncating and regenerating LLM Chain-of-Thought traces reveals a reliability signal for improved answer aggregation.
• When LLMs Agree, Are They Right? Auditing Self-Consistency and Cross-Model Agreement as Confidence Signals — Agreement among LLMs or within a model's samples is an unreliable, conditional proxy for correctness, not a standalone confidence signal.
• When Does Delegation Beat Majority? A Delegation-Based Aggregator for Multi-Sample LLM Inference — PPV beats majority voting in LLM inference by using delegation based on entropy and reasoning geometry.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.