When Is Next-Token Prediction Useful? Marginalization, Ergodicity, Mixture Identifiability, Local Sufficiency, RAG, Tools, and Programming
A new theoretical framework challenges the standard interpretation of language model training, arguing that next-token prediction alone cannot capture how LLMs actually generate text in real-world contexts. The paper distinguishes between the full conditional distribution (which includes latent circumstances like intent and context), the marginal text-only distribution, and what models actually learn from finite data. This distinction has direct implications for how practitioners should think about RAG, tool use, and code generation, where external constraints and non-textual conditioning are essential. The work suggests current training paradigms may be fundamentally incomplete for tasks requiring grounding beyond token sequences.
Modelwire context
ExplainerThe paper's most underreported contribution is the identifiability argument: it formalizes why models trained on marginal text distributions cannot, in principle, recover the latent conditioning signals that make outputs useful in grounded tasks, regardless of scale or data volume. This is a structural limit, not an engineering gap.
This connects directly to the 'Convergence Without Understanding' paper from the same day, which found that models develop similar representations but diverge on reasoning. That work identified a behavioral fragmentation problem; this paper offers a possible theoretical root cause, the training objective itself cannot encode the latent circumstances that drive consistent reasoning. The temporal failure modes piece on statutory QA also illustrates the practical downstream cost: RAG-augmented retrieval partially compensates for missing context, but this framework suggests that compensation is doing work the base training objective was never designed to handle.
Watch whether any major training framework (Hugging Face, DeepMind, or OpenAI) formally incorporates latent-variable conditioning into a published pretraining objective within the next 12 months. If they do not, this paper will likely remain a diagnostic tool rather than a prescriptive one.
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsLanguage models · RAG · Next-token prediction
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.