Researchers have proposed a new approach to reinforcement learning called Q-Value Expectile Disagreement (QED). The goal of QED is to reduce the variance in performance across different training runs by controlling behavioral divergence. This is achieved through a state-dependent temperature schedule that uses double-critic disagreement as a proxy for cross-run disagreement. The authors claim that QED reduces cross-run divergence by two orders of magnitude without sacrificing performance, resulting in lower return variance at modest sample-efficiency costs. --- Why it matters: This matters to engineers and researchers in AI because it addresses a major challenge in reinforcement learning: the high variance in performance across training runs. By reducing this variance, QED enables more reliable deployment of RL models in real-world domains. Source: https://arxiv.org/abs/2605.21214