Honeycomb: Constant-Size Scene Memory for Video World Models
Researchers from Harvard, MIT, NUS, and MIT-IBM Watson AI Lab propose Honeycomb, a video world model with a constant-size scene memory representation called HexMemory. HexMemory compresses scenes into six fixed-size 2D feature planes and updates the same memory block cyclically, preventing memory growth with video length. The method improves consistency when cameras revisit locations, addressing the problem of forgetting spatial details in long videos.
Coverage timeline
机器之心机器之心
让视频生成模型沿着相机轨迹不断探索一个场景，并不算最难。真正困难的是：当镜头绕了一圈再次回到原点，房间的布局、家具的位置和墙面的纹理，是否还与最初一样？ 现有视频世界模型可以生成视觉效果出色的短片。但随着生成过程不断延长，早期画面会逐渐离开模型有限的时间上下文。模型可能 “忘记” 已经见过的空间：沙发换了材质，门窗悄然移动，甚至房间的几何结构也发生变化。 一种直接的解决办法，是把过去观察到的内容全部保存下来。然而，无论保存的是 RGB 图像、三维点，还是扩散模型的潜特征，记忆都会随着视频长度持续增长。模型走得越远，需要保存和处理的历史就越多。 来自哈佛大学、麻省理工学院、新加坡国立大学与 MIT IBM Watson AI Lab 的研究团队提出 Honeycomb，试图回答一个更根本的问题：视频世界模型能否在持续吸收新场景信息的同时，让特征记忆始终保持固定大小？ Honeycomb 的答案是 HexMemory。它将场景压缩到六张固定尺寸的二维特征平面中，并通过 “读取、生成、写回” 的循环不断更新同一块记忆。无论视频生成多长，HexMemory 的特征存储量都不会随帧数继续增加。 图 1：闭环重访动态对比。从左到右依次为 Honeycomb、Spatia 与 LSM World。相机离开并返回起点后，Honeycomb 更好地保持了房间结构和物体外观。来源：Honeycomb 项目主页。 论文标题与作者信息。Jack Wei Lun Shi 与 Kaichen Zhou 为共同第一作者。来源：论文首页。 论文标题：Honeycomb: Constant-Size Scene Memory Representation for Video World Models 作者：Jack Wei Lun Shi (Co-first)、Kaichen Zhou (Co-first)、Haoyu Chen、Yufeng Weng、Keane Ong、Ruojin Cai、Hang Hua、Justin K.W. Yeoh、Mengyu Wang 论文：https://arxiv.org/abs/2609.37690 项目主页：jackswl.github.io/honeycomb 代码：github.com/kaichen-z/honeycomb 为什么长视频需要一块不会膨胀
Related stories
Apple and Duke propose LensVLM, compressing long context into images with on-demand zoom
Apple's foundation model team and Duke University introduced LensVLM, a method that compresses long context into images and lets the model zoom into relevant pages on demand. In seven text QA benchmarks, it achieved 68.9% average accuracy with 4.3x effective compression, close to the 72.4% of full-text reference. The work was accepted at NeurIPS 2026, and a public 9B demo is available.
1 reports · 1 sources
Oxford, Monash, Bristol Researchers Propose Oneira Interactive Video World Model
Researchers from Oxford, Monash, and Bristol universities propose Oneira, an interactive video world model that separates world state from visual rendering. A Coding Agent maintains a World State Table to record object positions and states, enabling precise, persistent object changes even when the camera moves away and back. The model addresses limitations of existing video world models that lack explicit state tracking, leading to wrong interactions or lost modifications.
Anthropic: Small AI models may self-replicate on IoT devices within months
Anthropic researchers report that small AI models may be months away from self-replicating on commercial IoT devices like smart ovens and off-brand Android TV boxes. In tests, a 9B model ran inference on 8 of 9 hardware types, and GLM-5.2 autonomously spread an 8B model to all 8 nodes in 4 of 5 attempts. The researchers predict models under 30B could achieve end-to-end replication by the end of 2026.
LAMA Framework Auctions Ads Token-by-Token During LLM Generation
Researchers from Renmin University and Stanford propose LAMA, a large-model-native advertising framework that auctions ads per generated token, integrating ad allocation into the generation process. This shifts from traditional pre-defined ad slots to dynamic allocation where each token influences both content continuation and ad exposure.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe