Cambridge paper trains LLMs to ask informative questions via information-gain rewards
_akhaliq · x · 2026-10-30
A University of Cambridge paper by Anita Kriz, Antonin Berthon, Nicolás Astorga and Mihaela van der Schaar tackles a known LLM weakness: adaptively gathering missing information in multi-turn settings.
- Method: amortize sequential Bayesian experimental design (BED) into an LLM by explicitly training the reasoning that guides action selection with information-gain rewards; the experimental generative model is specified independently of the LLM's own beliefs.
- Results: on Twenty Questions, the learned policies beat both inference-time and training-time BED baselines at substantially lower compute and generalize to unseen categories.
- Transfer: training transfers positively to Medi-Q, a real-world clinical dialogue benchmark, suggesting information gathering can be learned as a generalizable reasoning skill.
Related event: Cambridge Trains LLMs with Information-Gain Rewards for Experimental Design(2 posts)→
More from Research
- SAT solver proves pigeonhole UNSAT in 33,423 steps, verified by VeriPB — TinfoilTricorn · 2026-10-30
- Datology AI builds synthetic pretraining data with measurable knowledge-cutoff guarantees for LLM backtesting — josh_wills · 2026-10-30
- Second-hand electronics price forecasting gets a systematic benchmark: deep learning vs statistical models — PtrPomorski · 2026-10-30
- Training a single random neuron turns Qwen3-8B into an owl, matching full SFT — maxsloef · 2026-10-30
- Mathematicians link distinct knots to formulas developed by Ramanujan, per Scientific American — Zulfikar_Ramzan · 2026-10-30
- Philosopher argues AI could usher in a golden age of mathematics—if we redefine mathematical progress — Dr_Atoosa · 2026-10-30