LC-ERD: Mining Latent Logic for Self-Evolving Reasoning via Consistency-Regulated Reward Decomposition
The paper introduces LC-ERD, a framework designed to enhance self-evolving reasoning in Large Language Models. It addresses challenges such as label noise, coarse-grained supervision, and distributional collapse. The proposed method aims to improve logical consistency and accuracy in reasoning patterns through a novel reward decomposition approach.
- ▪LC-ERD stands for Logic-Consistent Endogenous Reward Decomposition.
- ▪The framework aims to mine latent logic to improve reasoning in Large Language Models.
- ▪It addresses issues like label noise and coarse-grained supervision that hinder effective self-alignment.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.24005 (cs) [Submitted on 19 May 2026] Title:LC-ERD: Mining Latent Logic for Self-Evolving Reasoning via Consistency-Regulated Reward Decomposition Authors:Yanyu Chen, Jiyue Jiang, Dianzhi Yu, Zheng Wu, Jiahong Liu, Jiaming Han, Xiao Guo, Jinhu Qi, Yu Li, Yifei Zhang, Irwin King View a PDF of the paper titled LC-ERD: Mining Latent Logic for Self-Evolving Reasoning via Consistency-Regulated Reward Decomposition, by Yanyu Chen and 10 other authors View PDF HTML (experimental) Abstract:The evolution of Large Language Model (LLM) reasoning is bottlenecked by the scarcity of high-quality process data.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
