LC-ERD: A Framework for Self-Evolving LLM Reasoning via Latent Logic Mining
A new framework called LC-ERD (Logic-Consistent Endogenous Reward Decomposition) has been developed by researchers to overcome limitations in reasoning within Large Language Models (LLMs) due to a lack of high-quality process data. This approach addresses three main issues: mimetic bias, which leads to label noise by favoring statistical likelihood over logical accuracy; coarse-grained supervision that provides insufficient detailed guidance; and distributional collapse, where signals do not generalize effectively. LC-ERD conceptualizes self-alignment as mining latent structures, creating a Variational Logic Potential by consolidating consensus from the model’s Latent Logic Expertise (LLE) to refine the reasoning process. This method seeks to facilitate self-improving reasoning independent of external supervision. The research is published on arXiv with ID 2605.24005.
Key facts
- LC-ERD stands for Logic-Consistent Endogenous Reward Decomposition.
- It addresses three challenges: label noise via mimetic bias, coarse-grained supervision, and distributional collapse.
- The framework uses Latent Logic Expertise (LLE) to aggregate consensus.
- It derives a Variational Logic Potential to denoise the reasoning manifold.
- The approach aims for self-alignment without external supervision.
- The paper is published on arXiv with ID 2605.24005.
- The method is designed for Large Language Models (LLMs).
- It frames self-alignment as latent structure mining.
Entities
Institutions
- arXiv