Do We Need Frontier Models to Verify Mathematical Proofs?

TL;DR

This paper evaluates the proof verification capabilities of various language models, demonstrating that smaller models can match frontier models' performance with specialized prompts, improving reliability and accuracy.

Contribution

It reveals that smaller open-source models can verify proofs as effectively as frontier models when guided by optimized prompts, challenging assumptions about model size requirements.

Findings

Smaller models are up to 10% less accurate than frontier models.

Specialized prompts improve smaller models' accuracy by up to 9.1%.

Ensemble prompts enable smaller models to match frontier model performance.

Abstract

Advances in training, post-training, and inference-time methods have enabled frontier reasoning models to win gold medals in math competitions and settle challenging open problems. Gaining trust in the responses of these models requires that natural language proofs be checked for errors. LLM judges are increasingly being adopted to meet the growing demand for evaluating such proofs. While verification is considered easier than generation, what model capability does reliable verification actually require? We systematically evaluate four open-source and two frontier LLMs on datasets of human-graded natural language proofs of competition-level problems. We consider two key metrics: verifier accuracy and self-consistency (the rate of agreement across repeated judgments on the same proof). We observe that smaller open-source models are only up to ~10% behind frontier models in accuracy but they are up to ~25% more inconsistent. Furthermore, we see that verifier accuracy is sensitive to prompt choice across all models. We then demonstrate that the smaller models, in fact, do possess the mathematical capabilities to verify proofs at the level of frontier models, but they struggle to reliably elicit these capabilities with general judging prompts. Through an LLM-guided prompt search, we synthesize an ensemble of specialized prompts that overcome the specific failure modes of smaller models, boosting their performance by up to 9.1% in accuracy and 15.9% in self-consistency. These gains are realized across models and datasets, allowing models like Qwen3.5-35B to perform on par with frontier models such as Gemini 3.1 Pro for proof verification.