New bounds show large language models can generalize beyond training data.
problem Generalization of large language models beyond training data.
method Compression bound derivation using prediction smoothing and SubLoRA.
result Large language models can discover generalizable regularities.