A new training method teaches AI recommendation chatbots to ask questions that actually narrow down what you want, instead of just filling airtime.
Researchers propose fine-tuning large language models used as conversational recommender systems by rewarding each question based on how much it reduces the model's uncertainty about what to recommend, measured as entropy over its recommendation list. That reward doesn't require knowing the "correct" answer in advance, which matters because real conversations rarely come with ground truth. The team fine-tuned models using both supervised fine-tuning and direct preference optimization, then tested the approach on two standard conversational recommendation datasets, INSPIRED and ReDial. Both training methods improved recommendation quality and cut the number of turns needed to reach a good suggestion.
Most conversational recommenders today either follow scripted question templates or get graded by a second LLM judging how "interactive" a conversation feels, a proxy that never confirms the questions were actually useful. Tying the reward directly to uncertainty reduction gives these systems a real signal for efficient information-gathering, the kind of thing that determines whether a chatbot feels genuinely helpful or just talkative.
It's a fix for a specific, boring problem: bots that ask too many pointless questions before recommending something you didn't want anyway. Whether it holds up outside the INSPIRED and ReDial benchmarks, on messier real-world catalogs, is the next test.