AI Paper Atlas

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

ICLR 2026 · Large language models › Agents, code and math · approach: LLMs for optimization and heuristics

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 changedWhereMeasuredEffect
use reflective generate-critique-refine paradigm instead of one-pass generation (ablation)ReForm-8B · PutnamBenchsemantic consistency: 40.8 -> 57.3Improved
use reflective paradigm instead of one-pass (ablation)ReForm-8B · AIME2025semantic consistency: 16.7 -> 46.7Improved
use reflective paradigm instead of one-pass (ablation)ReForm-8B · miniF2Fsemantic consistency: 82.7 -> 87.7Improved
add bounded clipping of prospective returns (ablation)PBSO · AIME2025semantic consistency: 26.7 -> 46.7Improved
add auxiliary critique reward r_aux (ablation)PBSO · PutnamBenchsemantic consistency: 52.1 -> 57.3Improved
add auxiliary critique reward r_aux (ablation)PBSO · miniF2F and ProofNetsemantic consistency: 87.7 -> 87.7 (miniF2F), 65.6 -> 65.6 (ProofNet)No clear effect
add RL training (PBSO) after SFT (ablation)ReForm-8B · ProofNetsemantic consistency: 62.3 -> 65.6Improved
scale ReForm from 8B to 32BReForm · AIME2025semantic consistency: 46.7 -> 66.7Improved
scale ReForm from 8B to 32BReForm · miniF2Fsyntactic correctness: 98.4 -> 97.1No clear effect
evaluate with CriticLean-14B judge instead of Qwen3-235BReForm-8B · average of 4 benchmarkssemantic consistency: 64.3 -> 70.3No clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy