From Tokens to States: LLMs as a Special Case of World Models and the Continuous Path Beyond

Abstract: The AI community has framed the relationship between LLMs and world models as a dichotomy: LLMs predict tokens; world models simulate reality. Yann LeCun argues in 2022 that reaching general intelligence requires abandoning autoregressive token prediction in favour of latent-space architectures. This framing is unnecessarily binary. Two claims will be defended. First, LLMs are a degenerate special case of world models: the state space is the set of all token sequences, the only action is appending one token, and world models are therefore a strict generalisation of LLMs, not a replacement. Second, there is a natural continuous spectrum from NTP to JEPA, with multi-token prediction, future-summary prediction, and next-latent prediction as intermediate stations already populated by current research. Moving along this spectrum relaxes the LLM constraints one by one. It also progressively surrenders the two practical advantages that make LLMs trainable at scale: internet-scale self-supervised data, and a transformer architecture co-designed for discrete token prediction. Both are examined as open research questions: the data question (the cliff from self-supervised text to instrumented action-labelled environments) and the architecture question (whether the transformer generalises to continuous-state prediction, or whether a new primitive is needed).

• The paper formally embeds autoregressive LLMs within the broader world-model framework, arguing that their hidden states can encode structured information about environments even when tokens serve as the interface.
• The paper maps a continuous progression from next-token prediction through multi-token, future-summary, and next-latent prediction to JEPA, showing how each step expands state modeling while weakening LLMs’ data or architectural advantages.
• The paper identifies a final data and architecture cliff for action-conditioned continuous world models, highlighting the need for efficient environment data and transformer alternatives suited to continuous sequential prediction.

This opinion paper by Paul Dubois argues against the prevailing dichotomy in which LLMs "predict tokens" while world models "simulate reality." The paper advances two claims. First, every autoregressive LLM can be formally embedded in the world-model formalism, making world models a strict generalisation of LLMs rather than a replacement. Second, there is a continuous spectrum from next-token prediction (NTP) to JEPA, with intermediate architectures already populated by current research, along which the constraints of the LLM special case are relaxed one at a time. The central tension the paper identifies is that each relaxation also erodes one of the two practical advantages that make LLMs trainable at scale: internet-scale self-supervised data and a transformer architecture co-designed for discrete token prediction.

Formal containment of LLMs within world models

The paper adopts the standard world-model tuple $(S, A, T, \rho_0)$: a state space, action space, transition function, and initial state distribution, with planning proceeding by simulating trajectories through iterative application of $T$ (Ha et al., 2018). An autoregressive LLM is then shown to instantiate this tuple with maximally restrictive choices:

Component	General world model	LLM
State space $S$	Arbitrary (continuous, multimodal, latent)	Finite token sequences $V^*$
Action space $A$	Semantic agent actions	Single next token
Transition $T$	Learned environment dynamics	Deterministic append
Supervision	Paired (state, action, next state) tuples	NTP on raw text
Architecture	Open (SSMs, diffusion nets)	Transformer

This yields the containment claim $\text{LLMs} \subset \text{World Models}$. The paper notes a structural subtlety: in the LLM instantiation the transition is trivial (append), so all predictive content resides in the policy $\pi_\theta$, which simultaneously plays the roles of world simulator and agent. This conflation is presented as a defining, not incidental, feature of the LLM special case.

The empirical support for treating the containment as more than a formal exercise comes from mechanistic interpretability. OthelloGPT, trained only on move tokens, encodes the full board state linearly in its hidden activations at greater than 99% accuracy (Li et al., 2022, Nanda et al., 2023); chess-playing LMs replicate this (Karvonen, 2024); and Llama-2 encodes linear representations of geographic space and calendar time (Gurnee et al., 2023). The paper draws a pointed inference: the world model is in the activations, not the tokens — tokens are the interface (the action space $A = V$), while the latent world state lives in the residual stream. Relatedly, hidden states at prompt level are shown to encode global attributes of the entire future response (Dong et al., 10 Feb 2025), indicating internal representations richer than the token-level objective strictly requires.

The spectrum from NTP to JEPA

The second claim organises existing architectures as stations along a spectrum, each step relaxing exactly one LLM constraint. The transitions are:

• NTP to multi-token prediction (MTP): relaxes the "one token per step" action granularity. Predicting $k$ tokens via independent heads improves reasoning and code performance (+15% on MBPP), and has been adopted in DeepSeek-V3 (Gloeckle et al., 2024, DeepSeek-AI et al., 2024). The paper characterises this step as a near-free upgrade: same data, same architecture, only $T$0 extra output heads. Theoretical work attributes the benefit to representational contractivity driving convergence toward internal belief states (Zhong et al., 7 Apr 2026).
• MTP to future summary prediction: decouples the prediction target from the token space by training an auxiliary head on compressed representations of the long-term future, improving maths and reasoning at 3B–8B scale while retaining internet text and an unchanged architecture (Mahajan et al., 16 Oct 2025).
• Future summary to next-latent prediction: moves the target fully out of token space. NextLat trains a transformer to predict its own next hidden state, improving planning and inference speed by up to 3.3× (Teoh et al., 8 Nov 2025). Critically, supervision comes from the model's own hidden states, so internet-scale corpora remain usable. However, the architecture fit weakens: predicting continuous latent vectors requires diffusion-style heads, and the transformer's inductive bias is no longer perfectly matched.
• Next-latent to JEPA: the point where both practical advantages collapse simultaneously. JEPA predicts action-conditioned latent future states [LeCun 2022], requiring paired (observation, action, next observation) data from instrumented environments — roughly $T$1 samples versus $T$2 tokens of text. The paper observes that existing JEPA models (I-JEPA, V-JEPA 2) sidestep this by re-discretising inputs into patches, effectively retreating toward the discrete-token end of the spectrum.

The data and architecture questions

The paper's analysis of the spectrum's costs is its most substantive contribution. It distinguishes a data cliff from an architecture question. The data cliff occurs only at the final step: the first four stations all train on internet-scale text, directly or via self-generated supervision. The bottleneck therefore shifts along the spectrum — compute at the LLM end, data collection and environment instrumentation at the JEPA end. The paper argues this asymmetry is underappreciated in debates that frame the LLM–world-model gap as purely architectural.

The architecture question is whether the transformer generalises to continuous-state, action-conditioned prediction. The paper documents a recurring workaround: re-discretisation, from ViT image patches (Dosovitskiy et al., 2020) to V-JEPA 2's video patch tokens. Its hypothesis is that the field lacks an architectural primitive as well-matched to continuous sequential prediction as the transformer is to discrete tokens — a state it analogises to NLP before 2017. The two gaps interact: a better continuous-state architecture could reduce the need for action-labelled data by learning more efficiently from self-supervised signal, so progress likely requires simultaneous advances on both fronts. This is stated as a hypothesis, not a result.

LeCun's critique revisited

The paper engages directly with LeCun's position paper, which argues that token prediction cannot yield the structured representations needed for planning and that JEPA constitutes a different paradigm. Both halves are contested. Against the first, the paper argues the critique conflates interface with representation: the interpretability evidence shows world-state structure in hidden activations, so "tokens, not states" mischaracterises what LLMs internally compute. Against the second, the paper reframes chain-of-thought and extended reasoning as movement along the spectrum's granularity axis — intermediate reasoning tokens increase the effective planning horizon within the token-state world model, and even without CoT the hidden state encodes a rough full-response plan (Dong et al., 10 Feb 2025). JEPA is thereby repositioned as the far end of a single spectrum rather than a separate paradigm. The paper also cites empirical work confirming that LLMs can function as text-based world models at sufficient scale (Li et al., 21 Dec 2025), and related work deriving STRIPS-style world models from NTP (Núñez-Molina et al., 16 Sep 2025).

Limitations and open questions

The paper is explicitly an opinion piece, and several of its commitments are assumptions rather than demonstrated results. The containment claim is formal but the interpretation of interpretability evidence — that linearly decodable board or spatial structure constitutes a "world model" in the planning-relevant sense — is contested elsewhere in the literature (Gupta et al., 15 Nov 2025). The claim that MTP is a "near-free upgrade" rests on benchmark gains at specific scales and one theoretical account (Zhong et al., 7 Apr 2026) whose generality is untested. The spectrum itself is a conceptual organisation: the intermediate stations were developed independently, and the paper provides no evidence that they compose or that each is a natural initialisation for the next, though it asserts this incremental-transition property. Most significantly, the paper leaves its two titular questions open: whether the transformer generalises to truly continuous action-conditioned dynamics, and whether an architectural primitive for continuous sequential prediction exists that matches the transformer's fit to text. It also does not quantify the data cliff beyond order-of-magnitude comparison, nor demonstrate that instrumented-environment data scarcity is binding given possible self-supervised workarounds beyond V-JEPA 2's implicit-action approximation.

Conclusion

The paper replaces the LLM-versus-world-model dichotomy with a formal containment claim and a spectrum of intermediate prediction objectives, arguing that the path from LLMs to action-conditioned latent world models is a progressive relaxation of five constraints rather than a paradigm break. Its practical prescription is that intermediate stations — MTP through Next-Latent — are attractive because they improve world-model capacity while preserving both LLM scaling advantages, whereas the JEPA step forfeits both simultaneously. The destination, on the paper's account, is already specified by LeCun; the open questions are what data and what architecture will support the final transition.