Alaya-EVOKE: Interactive world model with external memory

Breakthrough in interactive world models

Alaya-EVOKE just became the #1 trending paper on HuggingFace on August 14, 2026 with 107 upvotes, marking a significant advance in the field of interactive world models. This research addresses two core challenges that previous world models have always faced: limited memory and high latency.

The problem with current world models

Modern interactive world models need to satisfy three conflicting requirements: persistent memory, responsive interaction, and long-horizon generation. These requirements place contradictory demands on the model.

Maintaining history in the denoiser context or key-value cache incurs growing cost, forcing a trade-off between session length and retained memory. Meanwhile, low-latency interaction relies on few-step generation, but its capabilities are bounded by the teacher model. This is a paradox that no solution before EVOKE has fundamentally solved.

EVOKE’s breakthrough architecture

EVOKE addresses both limitations by externalizing persistent world state and redesigning the teacher for long-horizon interactive generation. Scene geometry is maintained in an external world state bank, indexed by camera, from which only view-relevant information is retrieved, keeping the denoiser context bounded as the session grows.

Rather than treating the teacher as a fixed generator, EVOKE designs it for long-horizon supervision. Its sparse attention combines chunk-wise grouping, retrieval of selected distant frames, and linear-attention global state, yielding linear growth in memory and compute while enabling supervision over long horizons.

Impressive results

With bounded context and recurrent external memory, EVOKE supports open-ended, continuously evolving generation. On a single H200 GPU at 384×640 resolution, each 1.5-second chunk is generated in 2.11 seconds - an impressive speed for an interactive world model.

As a three-step world model, EVOKE achieves state-of-the-art performance on WBench while remaining competitive on VBench-Long and VBench-2.0. The distribution-matching objective, applied under self-forced rollouts, transfers both capabilities to a three-step student that uses no classifier-free guidance, improving resistance to long-term drift while preserving responsive conditioning.

Practical applications

This architecture opens up many interesting applications:

Robotics and autonomous systems: EVOKE can be used to create simulation environments for robot training, where robots can interact with the virtual world continuously without being limited by context windows.

Gaming and VR: The ability to generate open-ended content with low latency makes EVOKE an ideal candidate for procedural content generation in games and virtual reality.

Video generation: The external memory architecture allows generating significantly longer videos than previous methods, while still maintaining consistency and responsiveness.

Why is this paper trending?

This paper attracted attention because it solved a fundamental problem that the AI community has struggled with for years: how to create world models that can maintain state continuously without exploding computational cost. The method of an external memory bank with camera-indexed retrieval is a simple but effective idea, allowing the model to scale to much longer scenarios than before.

Additionally, achieving three-step generation (only 3 denoising steps) while maintaining high quality is a significant achievement, as it significantly reduces inference time - a crucial factor for real-time applications.

Authors and source

The paper was written by Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang, and Feng Zhao from AlayaLab. The code and model have been made public on HuggingFace at AlayaLab/Evoke.

You can read the full paper on arXiv:2608.13546.