A new paper on arXiv proposes a fine-tuning approach that makes language models more Bayesian when reasoning about hidden variables from small amounts of data. The authors argue that while LMs are increasingly used for tasks where Bayesian inference is the normatively correct solution, standard supervised fine-tuning does not necessarily instill this kind of probabilistic reasoning.
The proposed method fine-tunes the model with a Bayesian objective, encouraging it to align its predictions with posterior beliefs rather than just matching surface-level training targets. According to the abstract, this yields models that are "as Bayesian as their beliefs allow," meaning the model's behavior is constrained by how well its internal representations capture the relevant uncertainty.
The paper does not claim that the models become perfect Bayesian reasoners. Instead, the improvement is framed as a practical step toward better calibration and more robust inference in low-data settings. The authors position the work as a bridge between probabilistic modeling and the flexible reasoning capabilities of large language models.