> They just choose the most probabilistic next token

That does not imply that a model should hallucinate. A trivial counterexample is a small LLM trained up to 100% accuracy to output x mod 100 for any input x in the range 0-1000000 and "I don't know" for any other input that is not a number in that range. Such model does not hallucinate, even if it's still just a probabilistic autoreggressive next token predictor. In fact, this is a point argued in this paper

> Hallucinations are inevitable only for base models. Many have argued that hallucinations are inevitable (Jones, 2025; Leffer, 2024; Xu et al., 2024). However, a non-hallucinating model could be easily created, using a question-answer database and a calculator, which answers a fixed set of questions such as “What is the chemical symbol for gold?” and well-formed mathematical calculations such as “3 + 8”, and otherwise outputs IDK. Moreover, the error lower-bound of Corollary 1 implies that language models which do not err must not be calibrated, i.e., δ must be large. As our derivations show, calibration-and, hence, errors—is a natural consequence of the standard cross-entropy objective. Indeed, empirical studies (Fig. 2) show that base models are often found to be calibrated, in contrast to post-trained models which may deviate from cross-entropy in favor of reinforcement learning.