Google teaches LLMs Bayesian reasoning via distillation

This is one of the more practically grounded LLM reasoning papers in recent memory — it directly attacks the multi-turn stagnation problem that anyone building long-running agents has run into.

• –Off-the-shelf models including Gemini-1.5 Pro, GPT-4.1 Mini, and Llama-3-70B all showed near-zero improvement after the first interaction round; Bayesian-taught models kept improving across five rounds
• –The distillation target is a symbolic Bayesian model — not human-labeled data — which makes the supervision signal cheap and principled
• –Crucially, models trained only on flight recommendations transferred to hotel bookings and real web shopping, suggesting the method teaches a generalizable inference skill, not a domain trick
• –Published in Nature Communications, not just a preprint — unusually high bar for applied ML work
• –The unresolved question is whether SFT is the right training objective here; some researchers argue RL would better approximate probabilistic inference