Large language models (LLMs) are the world’s best imitators, but when it comes to the cold, hard logic of updating beliefs based on new evidence, they are surprisingly stubborn. A team of Google researchers argue that the current generation of AI agents fall far short of ‘probabilistic reasoning’ – the ability to maintain and update ‘world models’ as new information arrives.
Solution? Stop trying to give them the right answer and start teaching them to guess like a mathematician.
The Problem: The ‘One-and-Done’ Plateau
While LLMs like Gemini-1.5 Pro and GPT-4.1 Mini can write code or summarize emails, they struggle as interactive agents. Imagine a flight booking assistant: it has to guess your preferences (price vs. duration) by seeing which flights you choose over several rounds.
The research team found that off-the-shelf LLMs—including heavyweights like Llama-3-70B and Quen-2.5-32B—showed ‘little or no improvement’ after the first round of negotiations. Whereas the ‘Bayesian assistant’ (a symbolic model using Bayes’ rule) becomes more accurate with each data point, standard LLMs freeze almost immediately, failing to adapt their internal ‘beliefs’ to the user’s specific reward function.
Meet Bayesian learning
The research team introduced a technique called bayesian learning. Instead of fine-tuning a model on (what they call) ‘correct’ data oracle teacher), they fixed it to mimic bayesian assistant-A model that explicitly uses Bayes’ rule to update the probability distribution over possible user preferences.
Here is the technical glitch:
- Work: Five rounds of flight recommendation negotiation. Flights are defined by features such as price, duration and stops.
- awards ceremony: A vector representing the user’s preferences (for example, a strong preference for low prices).
- last updated: After each round, the Bayesian Assistant updates this Back distribution based on East (initial assumptions) and Possibility (The probability that a user will choose a certain flight given a specific reward function).
by using Supervised Fine-Tuning (SFT) On these Bayesian interactions, the research team forced the LLM to adopt Process Under uncertainty lies the logic, not just the final outcome.
Why ‘educated guesses’ outweigh correct answers
The most counter-intuitive conclusion of the research is that bayesian learning consistently outperformed oracle tutorial.
In ‘Oracle Teaching’, the model is trained on a teacher who already knows what the user wants. In ‘Bayesian teaching’, the teacher is often Wrong In the early stages because it is still learning. However, those ‘educated guesses’ provide a much stronger learning signal. By watching the Bayesian assistant struggle with uncertainty and then update its beliefs after receiving feedback, the LLM learns the ‘skill’ of belief updating.
The results were clear: Bayesian-tuned models (such as Gemma-2-9b or Llama-3-8b) were not only more accurate, but also agreed with the ‘gold standard’ Bayesian strategy about 80% of the time – significantly higher than their original versions.
Generalization: beyond flights to web shopping
For devs, the ‘holy grail’ is a generalization. A model trained on flight data not only needs to be good at flights; this must be understood concept Learning from a user.
The research team tested their refined models:
- increased complexity: Moving from four flight facilities to eight.
- new domains: Hotel Recommendations.
- real world scenarios: A web shopping task using real products (title and description) from a simulated environment.
Even though the models were only fine-tuned on synthetic flight data, they successfully transferred those probabilistic reasoning skills to hotel booking and web shopping.. In fact, Bayesian LLMs even outperformed human participants in some rounds, because humans often deviate from standard reasoning parameters due to biases or inattention..
Neuro-symbolic bridge
This research highlights a unique strength of deep learning: the ability to transform a classic, symbolic model (Bayesian assistant) into a neural network (LLM)..
While symbolic models are great for simple, codified tasks, they are extremely difficult to build for ‘messy’ real-world domains like web shopping. By teaching LLM to copy Using the symbolic model strategy, it is possible to get the best of both worlds: the rigorous reasoning of a Bayesian and the flexible, natural-language understanding of a Transformer.
key takeaways
- LLMs struggle with belief updating: Off-the-shelf LLMs, including state-of-the-art models such as Gemini-1.5 Pro and GPT-4.1 Mini, fail to effectively update their beliefs as they receive new information, with performance often plateauing after a single interaction.
- Bayesian learning outperforms direct training: Teaching an LLM to mimic the ‘educated guesses’ and uncertainty of a standard Bayesian model is more effective than training it directly on the correct answers (oracle learning).
- Probabilistic skills generalize across domains: LLMs fine-tuned on simple synthetic tasks (e.g., flight recommendations) can successfully transfer their belief-updating skills to more complex, real-world scenarios such as web shopping and hotel recommendations.
- Neural models are more robust to human noise: While a purely symbolic Bayesian model is optimal for consistent simulated users, fine-tuned LLMs exhibit greater robustness when interacting with humans, whose preferences often deviate from their stated preferences due to noise or bias.
- Effective Distillation of Symbolic Strategies: Research proves that LLMs can learn to approximate complex symbolic reasoning strategies through supervised fine-tuning, allowing them to apply these strategies in domains too dirty or complex to be explicitly codified in classic symbolic models.
check out paper And technical details. Also, feel free to follow us Twitter And don’t forget to join us 120k+ ml subreddit and subscribe our newsletter. wait! Are you on Telegram? Now you can also connect with us on Telegram.
