Nature Machine Intelligence, Published online: 10 September 2026; doi:10.1038/s42256-026-01298-6

Shi et al. demonstrate a dual architecture approach for materials research, integrating two lightweight large language models for collaborative reasoning and scientific tool execution. The method achieves competitive performance while remaining affordable and locally deployable.

In this paper, we study the neural exploration strategy for contextual bandits. The dilemma of exploitation and exploration widely exists in real-world applications such as recommender systems, online advertising, and clinical trials. Contextual bandits provide principled methods to solve this dilemma, including two prevalent techniques: Thompson Sampling (TS), and Upper Confidence Bound (UCB). Neural contextual bandits have been studied to adapt to the non-linear reward function, combined with TS or UCB strategies for exploration. In this paper, we introduce, EE-Net, which is a novel framework to utilize another neural network to learn the potential gain of exploitation neural network for exploration, different from UCB-based and TS-based approaches that rely on the large-deviation-based statistical confidence bound. In addition, we provide an instance-based $\widetilde{\mathcal{O}}(\sqrt{T})$ regret upper bound for EE-Net with a new proof workflow. Empirically, we show that EE-Net outperforms related linear and neural contextual bandit baselines on real-world datasets.