Step-Controlled DPO: Leveraging Stepwise Error for Enhanced Mathematical Reasoning
Research paper by Zimu Lu and collaborators. Published 2024-06-30. Topics: Direct Preference Optimization, Step-Controlled DPO, negative samples, mathematical reasoning rationales, automatic stepwise error supervision.
Resource information
- Authors
- Zimu Lu, Aojun Zhou, Ke Wang, Houxing Ren, Weikang Shi, Junting Pan, Mingjie Zhan
- Research topics
- Direct Preference Optimization, Step-Controlled DPO, negative samples, mathematical reasoning rationales, automatic stepwise error supervision, SFT models, chain-of-thought solutions, credit assignment, InternLM2-20B, GSM8K, MATH
- Publication date
- 2024-06-30
- Publication
- arXiv
- Last record update
- Nov 3, 2026