Apple presents BED-LLM, a technique that teaches LLMs to ask the right questions at the right time

Researchers at Apple, in partnership with universities such as Oxford and City University of Hong Kong, presented BED-LLM (Bayesian Experimental Design with Large Language Models), an innovative approach that enables language models to collect information adaptively and strategically across multiple interactions. The proposal addresses a critical limitation of current LLMs: difficulty adjusting questions based on previous answers. In practice, BED-LLM treats each interaction as a sequential experiment. Rather than merely generating plausible questions, the model selects, at each turn, the question that maximizes expected information gain, using principles of Bayesian experimental design. This allows the system to reduce uncertainty in a measurable way, approximating the model's behavior to that of a rational agent that learns actively. The technical differentiator lies in the explicit use of the LLM's own uncertainties. The method builds a probabilistic model from the language model's predictive distributions and avoids assuming deterministic responses. As a result, ambiguous, irrelevant, or low-information questions tend to be discarded automatically, as they offer low informational gain. In practical tests, BED-LLM showed significant gains. In the classic "20 Questions" game, the method multiplied the success rate by almost six compared to direct prompts, even using relatively small models. Similar results were observed in preference inference scenarios, such as movie recommendations, indicating robustness even when the model and user hold divergent beliefs. This advance has direct implications for conversational agents, recommendation systems, assisted diagnosis, intelligent tutoring, and automation of complex tasks. By allowing LLMs to decide not only what to answer but what to ask next, BED-LLM represents an important step toward truly interactive, decision-oriented agents.