---
title: "Code World Model: Coding Agent as World Brain for Persistent World Evolution"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.25927v1"
permalink: "https://feed.astralor.com/items/itm_d3796f7ce60b690e19bc9e0d"
published: "2026-08-26T15:37:33.000Z"
score: 77
channel: "tech"
tags: ["MiniMax-H3","World Model","Proxy Representation","Coding Agents","Video Generation"]
---

# Code World Model: Coding Agent as World Brain for Persistent World Evolution

## AI commentary

> What I find most interesting is using code as persistent memory while letting the video model only render, solving the problem of world models being 'myopic'—guessing rules purely from pixels. For those working on open-ended simulation and embodied AI, this framework is worth a close read.

## At a glance

Existing video-based world models primarily learn dynamics from visual observations, which is hard to maintain persistent consequences and support coherent open-ended evolution. This paper introduces Code World Model, which separates world evolution from visual realization: a coding agent acts as the world brain, reasoning about events and consequences and generating executable code to maintain persistent state, while a proxy representation encodes spatiotemporal constraints into a proxy video that conditions a video model to render high-fidelity observations. After fine-tuning on gameplay data, MiniMax-H3 follows proxy-based spatiotemporal specifications, demonstrating the potential of combining code for persistent evolution with video models for flexible realization.

## Key points

- Proposes Code World Model framework where a coding agent handles persistent world state and rule-consistent evolution
- Introduces proxy representation linking executable state to video generation, compiling spatiotemporal constraints into proxy video
- Builds data pipelines for aligned proxy-observation pairs from gameplay and real videos, fine-tuning MiniMax-H3 to follow proxy specifications
- Demonstrates a new path combining code for persistent evolution with video models for flexible realization

## Full text

World models aim to simulate how complex environments evolve under actions and events, yet existing video-based world models primarily learn dynamics from visual observations, which reveal outcomes rather than the underlying knowledge, rules, and mechanisms governing world evolution. This makes it difficult to maintain persistent consequences and support coherent, open-ended evolution. We introduce Code World Model, a framework that separates world evolution from visual realization by combining the reasoning and coding capabilities of language models with the generative priors of video models. A coding agent serves as the world brain, reasoning about events and their consequences and generating executable code to maintain persistent world state and perform rule-consistent evolution. To connect executable state with visual generation, we introduce a proxy representation that encodes frame-wise spatiotemporal constraints and is compiled into a proxy video, which conditions a video model to render high-fidelity visual observations. We further develop data pipelines for constructing aligned proxy-observation pairs from gameplay and real-world videos. After fine-tuning on paired gameplay data, MiniMax-H3 follows proxy-based spatiotemporal specifications from simple interactive worlds built by the coding agent while preserving rich visual details and dynamics. These results demonstrate the potential of combining code for persistent world evolution with video models for flexible visual realization, providing a new path toward open-ended world models.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_d3796f7ce60b690e19bc9e0d