Stochastic Autoregressive Learning

Abstract

Motivated by LLMs, which generate outputs by iteratively sampling from next-token distributions, we introduce a PAC-learning model for binary stochastic autoregressive learning. This generalizes the deterministic autoregressive learning framework of Joshi et al., COLT 2025. In our model, one fixed generator assigns a Bernoulli next-token distribution to every prompt string. Starting from an input prompt, a token is sampled and appended to the prompt; the same generator is then applied again to this expanded prompt; this procedure is repeated for $M$ steps. Three forms of supervision are considered: base one-step samples, chain-of-thought (CoT) samples that reveal full random trajectories of length $M$, and end-to-end (e2e) samples that reveal only the final token of length $M$ trajectories. For a generator class, we study the minimum number of samples $m_{base}(\epsilon ),m_{CoT}(\epsilon ),m_{e2e}(\epsilon )$, resp., required to learn the one-step probabilities in the base model, and the final-token probability in the CoT and e2e models, under squared loss error~$\epsilon$. We show that stochastic autoregressive learning fundamentally differs from the deterministic theory. At scale $\epsilon$, there is no universal comparison between the three learning tasks: both $m_{CoT}/m_{base}$ and $m_{e2e}/m_{CoT}$ can be made simultaneously arbitrarily larger than $M/\epsilon$, the natural analogue for the existing deterministic results. Nevertheless, after altering scales, for every class, CoT learning at scale $\epsilon$ is upper-bounded by base learning at scale $\epsilon /M^2$, whereas e2e learning at scale $\epsilon$ is upper-bounded, up to logarithmic factors, by $(M/\epsilon )m_{CoT}(\Theta (\epsilon ))$. These dependencies and scales are essentially tight. We complement these bounds by studying dimension $d$ logistic functions in our model.