arXiv:2610.29289v1 Announce Type: cross Abstract: Why large language models trained via log-loss minimization over massive and diverse datasets exhibit broad capabilities and generalizable skills is not yet fully understood. Grau-Moya et al. (ICML 2024) offer a conceptual explanation through the lens of meta-learning: if sequence-generating environments $\nu$ are drawn from a prior $w_{\nu}$, minimizing population log-loss recovers the corresponding Bayesian mixture predictor $\xi=\sum_\nu w_\nu\nu$. In the idealized limit in which $w$ is a universal prior, the optimal predictor corresponds to Solomonoff universal induction. Practical machine learning algorithms, however, generally produce only \emph{approximately} optimal predictors. This raises a natural question: \emph{Does the prediction performance remain robust under approximation}? In this paper, we first provide a tight bound on the worst-case effect of approximation on prediction error. We show any predictor $\rho$ satisfying $D_{KL}(\xi \|\rho)\leq \eta$ retains bounded cumulative-expected-squared prediction error. However, while the Bayesian mixture predictor $\xi$ has error $O(\log(1/w_\mu))$ in an environment $\mu$, we establish a tight $\Theta(\eta/w_\mu)$ bound on the excess error of $\rho$ relative to $\xi$. Thus, approximation can substantially degrade performance in environments assigned small prior probability. In sharp contrast, when environments are averaged according to prior $w$, we prove that the excess prediction error is only $O(\eta)$. Furthermore, we establish a distribution-shift theorem for dominated shifts from a training prior $w$ to a test prior $w'$, interpolating our average-case and worst-case guarantees. Finally, we extend our robustness analysis to a finite-sampling setting relevant to learning universal predictors with neural networks and Transformers, where $D_{KL}(\xi\|\rho)$ could be unbounded. [For the full abstract, see the paper.]