Researchers report lower-cost training for LLM information gathering

Researchers report in a preprint that training language models to select questions using information-gain rewards improved performance on Twenty Questions at substantially lower computational cost than the experimental-design baselines they tested. The trained policies also generalized to unseen categories. The method rewards reasoning that selects informative actions under an explicitly specified model of how experiments produce observations, separate from the language model's own beliefs. The researchers also report positive transfer to Medi-Q, a clinical dialogue benchmark. The demonstrated gains concern benchmark tasks, including clinical dialogue, rather than deployed clinical use.