用马尔可夫范畴理论解释语言模型生成机制与训练原理。

Autoregressive language models achieve remarkable performance, yet a unified theory explaining their internal mechanisms, how training shapes representations, and why these representations support complex behavior remains incomplete. We introduce an analytical framework that models the single-step generation process as a composition of information-processing stages using the language of Markov categories. This compositional perspective connects three aspects of language modeling that are often studied separately: the training objective, the geometry of the learned representation space, and practical model capabilities. First, our framework gives an information-theoretic rationale for parallel drafting methods such as speculative decoding by quantifying the information surplus a hidden state contains about future tokens beyond the immediate next one. Second, we clarify how the standard negative log-likelihood (NLL) objective learns not only a most likely next token, but also the data's intrinsic conditional uncertainty, formalized through categorical entropy. Our main spectral result is conditional: for a linear-softmax head with bounded output features, a calibrated quadratic upper-bound surrogate to NLL induces, after whitening or variance normalization, a generalized CCA/eigenproblem aligning representation directions with predictive prototypes. This gives a compositional lens for understanding how information flows through a model and how likelihood training can shape its internal geometry.