cvlab-kaist/worldkv

★ 17HTMLAudience · researcherComplexity · 5/5Setup · hard

Mindmap

mindmap
  root((worldkv))
    What it does
      KV cache compression
      World model memory
      Long sequence support
    Problem solved
      Cache grows over time
      Expensive long inference
      Memory bottleneck
    Target frameworks
      Lingbot-World-Fast
      Inspatio-World
      Matrix-Game-2.0
    Current status
      Paper published
      Code coming soon

mindmap root((worldkv)) What it does KV cache compression World model memory Long sequence support Problem solved Cache grows over time Expensive long inference Memory bottleneck Target frameworks Lingbot-World-Fast Inspatio-World Matrix-Game-2.0 Current status Paper published Code coming soon

Click or tap to explore — scroll the page freely

Things people build with this

USE CASE 1

Read the WorldKV paper to understand how retrieval-and-compression reduces memory cost in AI world models for video generation and game agents.

USE CASE 2

Use the pending code release to reproduce WorldKV results on Lingbot-World-Fast, Inspatio-World, or Matrix-Game-2.0 once the authors publish it.

USE CASE 3

Apply the WorldKV memory strategy to your own transformer-based world model to reduce inference cost on long sequences.

Getting it running

Difficulty · hard Time to first run · 1day+

Code not yet released, the repository is a placeholder while authors prepare the implementation.

In plain English

WorldKV is the official code repository for a research paper titled "WorldKV: Efficient World Memory with World Retrieval and Compression," authored by researchers at KAIST AI and Naver AI. The paper addresses a problem in AI world models, which are systems trained to simulate how a visual environment changes over time. These models power applications like video generation and game-playing agents that need to predict what comes next in a scene. As these models run, they accumulate memory in the form of a key-value cache, and that cache grows with every frame or step, making long-horizon generation increasingly expensive. The KV in WorldKV refers to this key-value cache. The paper proposes combining retrieval and compression to keep only the most relevant memory at each step rather than accumulating everything. The claim is that this makes world models more practical for longer sequences without sacrificing much quality. The repository is a placeholder at the time of writing. The README states that the authors are cleaning up the code and plan to release it in early June. Three specific implementations are listed as pending checkboxes, targeting three different world model frameworks: Lingbot-World-Fast, Inspatio-World, and Matrix-Game-2.0. None are checked off, and no setup, training, or inference instructions are present yet. If you are looking for runnable code, the project is not ready. The paper is the current deliverable, the code is described as coming soon.

Copy-paste prompts

Prompt 1

WorldKV proposes combining retrieval and compression to manage a growing key-value cache in a world model. Based on that concept, how would I implement a retrieval step that selects only the most relevant past key-value pairs during video generation?

Prompt 2

I want to apply WorldKV-style memory management to a transformer world model I am building. Explain what a key-value cache is in this context and how retrieval-based compression would work in a generation loop.

Prompt 3

The WorldKV paper targets three frameworks: Lingbot-World-Fast, Inspatio-World, and Matrix-Game-2.0. What do each of these world model frameworks do, and why would reducing their KV cache size matter for practical use?

Open on GitHub → Explain another repo

← cvlab-kaist on gitmyhub — every repo by this author, as a profile.

Verify against the repo before relying on details.