LayerRecall: ZJU & HKU fix long-video amnesia with selective memory routing

新智元 · wechat · 2026-09-06

Autoregressive video models forget identities and props when characters leave and return, because KV caches prioritize recent frames. ZJU/HKU propose LayerRecall, routing retrieved history K/V only into memory-sensitive DiT layers (what to retrieve + where to use), trained via Cross-Horizon Prediction Matching with a 384-frame teacher supervising a 32-frame student—only 1.65M trainable params on a frozen 5B backbone. Results: MemoBench Overall 0.548 (best baseline 0.531), MovieBench 0.578, VBench-Long parity, high-frequency flicker power ratio 0.60→0.38, 3.5s overhead. The router transfers zero-shot to LongLive and Self-Forcing. 6.6MB weights, training code and 100 eval prompts open-sourced.

Original post →

More from Multimodal

Multimodal channel →