FormalRewardBench benchmark evaluates LLM reward models for theorem proving

Researchers have introduced FormalRewardBench, a new benchmark designed to evaluate reward models used in formal theorem proving. This benchmark addresses the challenge of sparse credit assignment in reinforcement learning for theorem provers by enabling the comparison of reward models without extensive retraining. FormalRewardBench includes 250 preference pairs with various error injection strategies and has been used to test several large language models, revealing that frontier models perform best in evaluating proof quality. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT This benchmark aims to improve reward models for AI theorem provers, potentially leading to more capable AI systems in formal mathematics and complex reasoning tasks.

RANK_REASON The cluster describes a new academic paper introducing a benchmark for evaluating AI models in a specific domain. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• FormalRewardBench
• Lean 4
• Claude Opus 4.5
• CompassJudger-1-14B
• Qwen2.5-72B-Instruct
• DeepSeek-Prover-V2-7B

• paper
• safety