The field of large language models (LLMs) is rapidly evolving, with a focus on improving their reliability, trustworthiness, and ability to self-improve. Recent research has highlighted the importance of semantic calibration, uncertainty estimation, and coherence mechanisms in achieving these goals. Notably, studies have shown that LLMs can be semantically calibrated, allowing them to meaningfully assess confidence in their responses, and that probability-only approaches can be used for efficient uncertainty estimation. Additionally, research has explored the role of past and future context predictability in incremental language production, and proposed principled frameworks for self-improvement based on coherence mechanisms.

Noteworthy papers include: Trained on Tokens, Calibrated on Concepts, which provides a principled explanation of when and why semantic calibration emerges in LLMs. Probabilities Are All You Need, which proposes an efficient, training-free uncertainty estimation method that approximates predictive entropy using the responses' top-K probabilities. Coherence Mechanisms for Provable Self-Improvement, which establishes a formal framework for self-improvement based on coherence mechanisms and provides rigorous theoretical guarantees for monotonic improvement.