bayesian-experimental-design

论文提出 Halve-Then-Greedy（HTG）算法，用于固定预算下的自适应实验设计，以最大化推荐设计均值超过目标的后验概率，并证明了近最优的实例相关样本复杂度界。

This paper proposes amortizing sequential Bayesian experimental design into LLMs by training the reasoning that guides action selection with information-gain rewards. Trained on Twenty Questions, the resulting policies outperform training- and inference-time BED baselines at lower cost, generalize to unseen categories, and transfer to the Medi-Q clinical dialogue benchmark.

This paper formulates cognitive experiment design as a Bayesian Experimental Design problem, introducing an amortized framework that efficiently identifies maximally informative environments for inferring latent planning parameters, with validated performance on the Mouselab-MDP paradigm.