Research map / Large language models / Agents, code and math
ReForm: Reflective Autoformalization with Prospective Bounded Sequence Optimization
Guoxin Chen, Wu Jing, Xinjie Chen, Xin Zhao, Ruihua Song, Chengxi Li, Kai Fan, Dayiheng Liu, Minpeng Liao
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 |
|---|---|---|---|
| use reflective generate-critique-refine paradigm instead of one-pass generation (ablation) | ReForm-8B · PutnamBench | semantic consistency: 40.8 -> 57.3 | Improved |
| use reflective paradigm instead of one-pass (ablation) | ReForm-8B · AIME2025 | semantic consistency: 16.7 -> 46.7 | Improved |
| use reflective paradigm instead of one-pass (ablation) | ReForm-8B · miniF2F | semantic consistency: 82.7 -> 87.7 | Improved |
| add bounded clipping of prospective returns (ablation) | PBSO · AIME2025 | semantic consistency: 26.7 -> 46.7 | Improved |
| add auxiliary critique reward r_aux (ablation) | PBSO · PutnamBench | semantic consistency: 52.1 -> 57.3 | Improved |
| add auxiliary critique reward r_aux (ablation) | PBSO · miniF2F and ProofNet | semantic consistency: 87.7 -> 87.7 (miniF2F), 65.6 -> 65.6 (ProofNet) | No clear effect |
| add RL training (PBSO) after SFT (ablation) | ReForm-8B · ProofNet | semantic consistency: 62.3 -> 65.6 | Improved |
| scale ReForm from 8B to 32B | ReForm · AIME2025 | semantic consistency: 46.7 -> 66.7 | Improved |
| scale ReForm from 8B to 32B | ReForm · miniF2F | syntactic correctness: 98.4 -> 97.1 | No clear effect |
| evaluate with CriticLean-14B judge instead of Qwen3-235B | ReForm-8B · average of 4 benchmarks | semantic consistency: 64.3 -> 70.3 | No clear effect |
Closest papers by meaning
- Autoformalize Mathematical Statements by Symbolic Equivalence and Semantic Consistency (NeurIPS 2024 · similarity 0.92)
- FormalAlign: Automated Alignment Evaluation for Autoformalization (ICLR 2025 · similarity 0.90)
- ATLAS: Autoformalizing Theorems through Lifting, Augmentation, and Synthesis of Data (NeurIPS 2025 · similarity 0.89)
- Multi-language Diversity Benefits Autoformalization (NeurIPS 2024 · similarity 0.87)
- Autoformalizing Euclidean Geometry (ICML 2024 · similarity 0.87)
- Autonomous Evaluation of LLMs for Truth Maintenance and Reasoning Tasks (ICLR 2025 · similarity 0.87)
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.
