A research team has found a way to make language models dramatically better at asking useful questions in multi-turn conversations.
The technique, called Amortised Sequential Information Gathering (ASIG), bakes a statistical framework called Bayesian Experimental Design into the model's training rather than applying it at inference time. The team fine-tuned a 7-billion-parameter model using a multi-turn reinforcement learning method with a reward signal based on expected information gain. On the 20 Questions benchmark, ASIG more than doubled the base model's success rate. Inference cost dropped by over 25x compared to BED-LLM, a competing approach that runs the same Bayesian logic at query time instead of baking it into weights.
The efficiency gap is the real story. Inference-time reasoning methods have become a popular way to squeeze more capability out of existing models, but they carry a compounding cost every time the model takes a turn. Shifting that logic into training means you pay once and redeploy — a trade-off that matters a lot at production scale. The team also tested ASIG on MediQ, a medical diagnosis benchmark the model had never seen, and found the information-seeking behavior transferred, which is harder to fake than benchmark scores on familiar tasks.
The caveat is that this is a 7B model on controlled benchmarks, not a deployed clinical tool — the gap between "transfers out of distribution on a benchmark" and "works reliably in the wild" remains as wide as ever.