Critique of World Model: A Generative Latent Prediction Architecture for World Modeling

A new preprint proposes a Generative Latent Prediction (GLP) architecture for building general-purpose world models, critiquing existing approaches and outlining a hierarchical, self-supervised framework designed to simulate actionable possibilities of the real world [1]. The essay, authored by Mingkai Deng, defines the primary goal of a world model as "simulating all actionable possibilities of the real world for purposeful reasoning and acting" [1]. It surveys key design dimensions—data, representation, architecture, learning objective, and usage—and analyzes tradeoffs in current methods [1]. The work draws inspiration from the concept of "hypothetical thinking" in psychology literature and the science fiction classic Dune [1]. The paper was first submitted to arXiv on 7 July 2025, with revisions on 18 July 2025, 27 July 2025, and a fourth version on 16 June 2026 [1]. The proposed GLP architecture is built on stateful, hierarchical, multi-level, and mixed continuous/discrete representations [1]. It employs a generative and self-supervised learning framework [1]. According to the paper, rather than abandoning generative modeling to avoid signal variability, the GLP approach adopts hierarchical abstraction [2]. The architecture decomposes the problem across multiple layers of latent prediction, each specialized for different representational granularities, whether continuous perceptual features or discrete conceptual tokens [2]. This allows each layer to operate at an appropriate level of abstraction while remaining generative and predictive [2]. The GLP design consists of an Encoder and Decoder that together form a generative bottleneck grounding predictions in observable data, plus a Latent Reasoning Backbone composed of an enhanced large language model for discrete concept-based reasoning and a diffusion-based next-embedding predictor for continuous perceptual dynamics [3]. The paper argues that this encoder-world-model-decoder design leads to stronger supervision and more stable training dynamics than encoder-only approaches such as Joint Embedding Predictive Architecture (JEPA) [2]. The authors note that in JEPA, supervision occurs solely in the latent space rather than observation space, trading off challenges of pixel-level variability for the risk of indefinability, where predicted latents are not directly grounded in observable data [2]. The essay situates world modeling within the broader pursuit of artificial general intelligence (AGI), a field where companies such as OpenAI, Google DeepMind and Meta aim to create AI that can complete virtually any cognitive task at least as well as a human [5]. The paper outlines an outlook for a Physical, Agentic, and Nested (PAN) AGI system enabled by such a model [1]. The preprint has undergone four revisions, with the file size changing from 1,063 KB in the first three submissions to 1,007 KB in the fourth [1].