Research map / Large language models / Agents, code and math
VerifyBench: A Systematic Benchmark for Evaluating Reasoning Verifiers Across Domains
Xuzhao Li, Xuchen Li, Shiyu Hu, Yongzhen Guo, Wentao Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| allow 4k-token output instead of 8 tokens (ablation) | xVerify-9B-C · VerifyBench overall (full-CoT input) | recall: 84.76% -> 88.66% | Improved |
| allow 4k-token output instead of 8 tokens (ablation) | xVerify-9B-C · VerifyBench overall (full-CoT input) | accuracy: 90.96% -> 90.86% | No clear effect |
| allow 4k-token output instead of 8 tokens (ablation) | Qwen2.5-32B-Instruct · VerifyBench overall (full-CoT input) | recall: 83.58% -> 88.36% | Improved |
| allow 4k-token output instead of 8 tokens (ablation) | Qwen2.5-7B-Instruct · VerifyBench overall (full-CoT input) | recall: 93.74% -> 75.90% | Hurt |
| allow 4k-token output instead of 8 tokens (ablation) | Qwen2.5-14B-Instruct · VerifyBench overall (full-CoT input) | recall: 91.55% -> 80.21% | Hurt |
| allow 4k-token output instead of 8 tokens (ablation) | Qwen2.5-7B-Instruct · VerifyBench overall (full-CoT input) | accuracy: 83.52% -> 82.35% | No clear effect |
| allow 4k-token output instead of 8 tokens (ablation) | Qwen2.5-32B-Instruct · VerifyBench overall (full-CoT input) | accuracy: 85.34% -> 81.20% | Hurt |
| allow 4k-token output instead of 8 tokens (ablation) | xVerify-0.5B-I · VerifyBench overall (full-CoT input) | recall: 88.66% -> 84.76% | Hurt |
| allow 4k-token output instead of 8 tokens (ablation) | general-verify · VerifyBench overall (full-CoT input) | valid judgments: - -> 79.01% | Improved |
| use trained verifier xVerify-9B-C instead of general LLM Qwen3-14B (ablation) | full-CoT 4k-token output · VerifyBench overall | accuracy: 90.86% vs 91.11% | Mixed |
| use full-CoT input instead of boxed-only input (ablation) | xVerify-9B-C · VerifyBench overall (8-token output) | accuracy: 89.07% -> 90.96% | Improved |
Closest papers by meaning
- Reinforcing General Reasoning Without Verifiers (ICLR 2026 · similarity 0.93)
- VerifyBench: Benchmarking Reference-based Reward Systems for Large Language Models (ICLR 2026 · similarity 0.92)
- General-Reasoner: Advancing LLM Reasoning Across All Domains (NeurIPS 2025 · similarity 0.92)
- Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective (NeurIPS 2025 · similarity 0.91)
- Advancing LLM Reasoning with Natural Language and Numerical Feedback (ICML 2026 · similarity 0.91)
- Reinforcement Learning with Verifiable Rewards Implicitly Incentivizes Correct Reasoning in Base LLMs (ICLR 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
