Internal Coherence Maximization (ICM): A Label-Free, Unsupervised Training Framework for LLMs
Post-training methods for pre-trained language models (LMs) depend on human supervision through demonstrations or preference feedback to specify desired behaviors....

