History of World Model Evolution: Paradigm Shift from Symbolic Reasoning to Autonomous Agents

The article delves into the core position of world models in autonomous agent architectures, traces the three-generation paradigm evolution from explicit symbolic models to latent space neural dynamics and then to large language model-driven systems, and reveals the structural constraints of current systems due to the lack of physical foundations. The GitHub open-source project "World-Model-with-Autonomous-Agent" offers a systematic perspective to understand the deep meaning of related issues.

In cognitive science, world models are the core cognitive architecture for agents to understand the environment, predict the future, and make decisions. However, current mainstream LLM-based agent systems lack a solid physical foundation—they excel at handling symbols and text but have little knowledge of the causal mechanisms of the physical world. This "lack of physical foundation" is not a technical detail but a structural constraint, making pure statistical pattern matching insufficient to handle physical constraints, spatial relationships, and dynamic changes when interacting with the real world.

The project divides world model development into three main paradigms. First generation: Explicit symbolic world models—early AI systems relied on manually constructed symbolic representations and rule mappings (logical predicates, semantic networks, production rules). Advantages: high interpretability; limitations: knowledge acquisition bottleneck, difficulty handling uncertainty, fragility in open worlds. Second generation: Latent space neural world models—using encoder-decoder architectures to compress high-dimensional observations into low-dimensional latent spaces and learn state transition dynamics. Representatives: World Models (Ha & Schmidhuber,2018), PlaNet (Hafner et al.,2019), Dreamer series. Advantages: continuity and differentiability enabling gradient-based optimization; limitation: latent space is a "black box" with poor interpretability.

The cutting-edge direction combines large pre-trained language models with world modeling. Discrete semantic world models use tokenization to convert continuous experiences into discrete symbols while retaining neural network representation learning capabilities, aiming to fuse the interpretability of symbolic methods and the expressive power of neural methods. LLM-based agents like ReAct, Reflexion, and Generative Agents show great potential but raise a key question: do language models truly build internal world models or just perform complex pattern matching?

The project raises a critical question: Are modern language-based agents really building internal world models, or are they merely driven by large-scale statistical patterns? Empirical studies provide mixed evidence: on one hand, LLMs exhibit amazing common sense reasoning and causal inference abilities; on the other hand, they still have obvious defects in physical intuition, spatial reasoning, and long-term planning. This suggests current systems may have a form of "weak world model" but are far from human-level physical understanding.

The project conducts small-scale experimental verification by reconstructing simplified scenarios inspired by "Generative Agents". The experiments focus on core issues of representation evolution: how different generations of agents define "intelligence" and how their internal representations differ when solving the same tasks. By comparing symbolic systems, latent space models, and LLM-based systems in the same environment, researchers can better understand the pros and cons of each paradigm.

The project points out key future research directions: 1) Strengthening physical foundations—explicitly integrating physical engines, geometric reasoning, and causal inference mechanisms into agent architectures; 2) Unifying representations—exploring how to fuse the interpretability of symbols, continuity of latent spaces, and richness of semantic representations into a unified framework;3) Innovating evaluation methods—developing strict test benchmarks that distinguish "true understanding" from "pattern matching". The evolution from symbolic to neural to hybrid architectures is not a simple replacement but a process of mutual learning and integration.