Stateful Harness splits rule execution from video world-model rendering
Original title: State World Models: Execute the Rules, Generate the Pixels
On 50 counterfactual twin pairs, MiniMax-H3 rises from 32% to 92% rule compliance with the harness.
Stateful Harness keeps an evolving world state under specified rules, renders each event with a frozen video model, and feeds the clip back to update state. A preview from an editable 3D scene supplies spatial and camera guidance. WorldState-Bench has 50 twin pairs that share an initial frame and action but differ in hidden state. Under rule-only prompting the best of 11 backends succeeds on both cases in 32% of pairs; specifying endings lifts MiniMax-H3 to 64%, and the full harness reaches 92%.
- WorldState-Bench uses 50 counterfactual twin pairs with shared frames and actions.
- Best of 11 video backends meets both required endings in 32% of pairs.
- MiniMax-H3 rises from 32% to 64% when required endings are specified.
- The full harness reaches 92% with the same frozen MiniMax-H3 model.
- Products
- Stateful HarnessWorldState-Bench
- AI models
- MiniMax-H3
From the original article
1 min read at huggingface.co
Existing world models can generate action-conditioned video from a single image. Many focus on camera movement, allowing users to explore the generated scene. However, navigation alone does not ensure that an action follows persistent rules when its outcome depends on state that is invisible in the current frame.
By Chan Yan Ki, Yanran Zhang, Wenzhao Zheng · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.