Research map / Large language models / Agents, code and math
SCI-Verifier: Scientific Verifier with Thinking
Shenghe Zheng, Chenyu Huang, Fangchen Yu, Junchi Yao, Jingqi Ye, Tao Chen, Yun Luo, Ning Ding, LEI BAI, Ganqu Cui, Peng Ye
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 |
|---|---|---|---|
| enable chain-of-thought reasoning before the verdict in verifiers (ablation) | verifier models of different scales · VerifierBench and VerifyBench-hard | accuracy | Improved |
| train with reasoning mode instead of no chain-of-thought (ablation) | SCI-Verifier · SCI-VerifyBench and two other benchmarks | accuracy | Improved |
| add RL after SFT on the base model (ablation) | SCI-Verifier (4B and 8B) · SCI-VerifyBench and generalization benchmarks | accuracy | Improved |
| apply RL directly on the Base model without SFT warm-up (ablation) | SCI-Verifier · SCI-VerifyBench | accuracy | Hurt |
| distill complete CoT instead of short CoT in SFT (ablation) | SCI-Verifier SFT · SCI-VerifyBench | accuracy and output length | Hurt |
| train on the proposed filtered dataset instead of the 180K-sample RM dataset (ablation) | Qwen3-4B-Base with SFT and SFT+RL · three verification benchmarks | accuracy | Improved |
| use a larger model size | verifiers on SCI-VerifyBench · SCI-VerifyBench | accuracy | No clear effect |
| use reasoning models instead of chat models | general LLMs as verifiers · SCI-VerifyBench | accuracy | No clear effect |
| change the prompt from the default to the xVerify-style prompt (ablation) | SCI-Verifier-8B · SCI-VerifyBench | accuracy: 86.28 -> 85.50 | No clear effect |
| change the prompt from the default to the xVerify-style prompt (ablation) | Qwen3-30B-A3B-Instruct-2507 · VerifyBench-Hard | accuracy: 88.70 -> 75.40 | Hurt |
| add equivalence-transformed answers and targeted post-training | SCI-Verifier vs GPT-5 · equivalence-augmented test set | accuracy | Improved |
Closest papers by meaning
- SciReason: A Controllable Benchmark for Scientific Reasoning in LLMs (NeurIPS 2026 · similarity 0.90)
- SciInstruct: a Self-Reflective Instruction Annotated Dataset for Training Scientific Language Models (NeurIPS 2024 · similarity 0.89)
- Variation in Verification: Understanding Verification Dynamics in Large Language Models (ICLR 2026 · similarity 0.89)
- CURIE: Evaluating LLMs on Multitask Scientific Long-Context Understanding and Reasoning (ICLR 2025 · similarity 0.89)
- Reinforcing General Reasoning Without Verifiers (ICLR 2026 · similarity 0.88)
- SciBench: Evaluating College-Level Scientific Problem-Solving Abilities of Large Language Models (ICML 2024 · similarity 0.88)
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.
