Alaya-EVOKE takes the memory out of the model, and the cost curve goes flat
Interactive world models have a boring, fatal problem: the longer you interact, the more history there is, and if that history lives in the model's working memory then cost grows quadratically and the session dies. Alaya-EVOKE's answer is to stop storing history in the model at all. Paper is arXiv 2608.13546, submitted August 13, from Yuanyang Yin, Gongxuan Wang, Yifan Zhan, Chuanhao Li, Kaipeng Zhang and Feng Zhao. It's sitting at 107 upvotes on HuggingFace, top of today's board.
The split is clean. Scene geometry gets maintained externally as persistent state. Generation queries it and pulls back only what's relevant to the current view. Context stops bloating because the thing that would have bloated it isn't in the context. To make that work at training time they rebuilt the teacher for extended-horizon supervision with a sparse attention scheme that combines three things: chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state. Memory and compute then grow linearly instead of quadratically, which is the whole ballgame.
The performance numbers are not a research-only artifact. On a single H200 it generates 1.5-second video chunks in 2.11 seconds at 384x640 as a three-step model, no classifier-free guidance. That's near-interactive on one GPU. It takes state of the art on WBench and stays competitive on VBench-Long and VBench-2.0, so the architecture isn't buying speed by giving up quality.
Here's why this belongs in an agents publication even though it's a video paper. This is exactly the argument the agent memory people have been having for two years, arriving from a completely different direction and landing in the same place: do not put your history in the context window, put it in a store and retrieve against it. Video made the argument unavoidable because the history is enormous and the interactivity requirement is brutal, so the quadratic wall shows up in seconds instead of hours. The conclusion transfers unchanged.
And it pairs uncomfortably well with PlayWorld, which also went up today and finds that current world models fail at exactly persistent state evolution and spatial consistency over long interaction. One paper says put the state outside the model. The other measures how badly models do when it's inside. Read them together.
← Back to all articles
The split is clean. Scene geometry gets maintained externally as persistent state. Generation queries it and pulls back only what's relevant to the current view. Context stops bloating because the thing that would have bloated it isn't in the context. To make that work at training time they rebuilt the teacher for extended-horizon supervision with a sparse attention scheme that combines three things: chunk-wise grouping, retrieval of selected distant frames, and a linear-attention global state. Memory and compute then grow linearly instead of quadratically, which is the whole ballgame.
The performance numbers are not a research-only artifact. On a single H200 it generates 1.5-second video chunks in 2.11 seconds at 384x640 as a three-step model, no classifier-free guidance. That's near-interactive on one GPU. It takes state of the art on WBench and stays competitive on VBench-Long and VBench-2.0, so the architecture isn't buying speed by giving up quality.
Here's why this belongs in an agents publication even though it's a video paper. This is exactly the argument the agent memory people have been having for two years, arriving from a completely different direction and landing in the same place: do not put your history in the context window, put it in a store and retrieve against it. Video made the argument unavoidable because the history is enormous and the interactivity requirement is brutal, so the quadratic wall shows up in seconds instead of hours. The conclusion transfers unchanged.
And it pairs uncomfortably well with PlayWorld, which also went up today and finds that current world models fail at exactly persistent state evolution and spatial consistency over long interaction. One paper says put the state outside the model. The other measures how badly models do when it's inside. Read them together.
Comments