Internal Coherence Maximization (ICM): A Label-Free, Unsupervised Training Framework for LLMs

Recent advancements in large language model (LLM) training have introduced the Internal Coherence Maximization (ICM) framework, a label-free, unsupervised post-training method designed to enhance model coherence without relying on human supervision or preference feedback. Traditional post-training techniques depend heavily on human demonstrations or feedback, which become unreliable as task complexity increases, leading to issues such as reward hacking and the exploitation of feedback flaws. ICM addresses these limitations by leveraging the model's internal logical consistency to improve performance, thereby reducing dependence on potentially flawed human supervision signals. This development is significant because it taps into the latent