**Abstract**
Apparently rational behaviors of autoregressive LLMs, such as in-context learning, have been attributed to implicit Bayesian inference: since training data is best explained as a mixture, the optimal next-token-predictor learns to implicitly infer latent concepts and completes prompts consistently with Bayesian inference. Although it is optimal in-distribution, Bayesian inference is generally suboptimal on out-of-distribution prompts due to model misspecification. As model behavior on OOD prompts is only weakly constrained by pretraining, it is not guaranteed that Bayesian behavior is extrapolated OOD. In this talk, we investigate with small-scale experiments the degree to which Bayesian inference remains a good model of LM behavior on OOD prompts. We first review related approaches from the literature. Then, focusing on small-scale compositional tasks - learning rules of formal languages - we show that Transformers can solve harder tasks than trained on, even in settings where the Bayes posterior is undefined. We highlight the role of task compositionality as a useful inductive bias in enabling models to learn more than the training data.
**Speaker Biography**
Szilvia is a second-year PhD student working with Professor Ferenc Huszár. Szilvia's research focuses on explaining emergent abilities of LLMs, such as in‑context learning and out‑of‑distribution generalisation, as well as related foundational questions in (algorithmic) information theory.