Researchers have developed a new approach to improving the reliability of large language models (LLMs) that operate over long periods. The method, called Comparison-Only Tiny Advisor (COTA), involves training a small comparator model to judge whether alternative actions are better than those proposed by the LLM. This comparator is trained using pairs of similar tasks and counterfactual branches. COTA has been tested on several benchmark tasks and outperforms existing methods, even when the auxiliary model is weaker than the LLM. The approach shows that runtime intervention can be effective without requiring a strong expert solver or critic. --- Why it matters: This matters to researchers in AI because it provides an alternative to traditional approaches to improving LLM reliability, which often rely on retraining or external experts. COTA's ability to improve performance even with weaker auxiliary models has significant implications for the practical deployment of LLMs. Source: https://arxiv.org/abs/2608.21027