Developments·Paper
A constant-size memory for video world models!
Researchers from MIT, Harvard, and NUS introduce Honeycomb, a video world model built on a fixed-size memory. The constant-size memory design targets long-horizon video generation without unbounded state growth.
9 engagement1 sourceSat, Oct 3, 2026, 02:40 AM
AI Overview
- MIT, Harvard, and NUS researchers introduce Honeycomb, a video world model built on a fixed-size memory (arXiv, 2026-10-03)
Related
Developments·Paper
Off-the-shelf VLMs outperform video embeddings in visual search
A new paper shows that off-the-shelf vision-language models (VLMs) beat specialized video embeddings for search over visual content. The finding suggests practitioners can leverage existing VLMs for visual search tasks without training custom video embedding models.
A new paper shows that off-the-shelf vision-language models (VLMs) beat specialized video embeddings for search over visual content. The finding suggests practitioners can leverage existing VLMs for visual search tasks without training custom video embedding models.
7 engagement1 sourcehackernewsTue, Aug 11, 2026, 06:52 PM
Developments·Paper
Dynamic Context Runtime: Bounded Attention over Unbounded History
A new runtime technique, Dynamic Context Runtime, enables bounded attention over unbounded history, allowing models to process arbitrarily long contexts without proportional compute growth. This could significantly reduce inference costs for long-context applications.
A new runtime technique, Dynamic Context Runtime, enables bounded attention over unbounded history, allowing models to process arbitrarily long contexts without proportional compute growth. This could significantly reduce inference costs for long-context applications.
1 engagement1 sourceredditThu, Aug 20, 2026, 05:12 PM
Developments·Paper
WorldCrafter: Video World Model with Implicit 3D-aware Memory for Consistent Long-Horizon Exploration
A new arXiv paper presents WorldCrafter, a video world model that learns a camera-queryable implicit 3D-aware memory to maintain consistency across long horizons and changing viewpoints. A memory encoder and pose-conditioned readout compress multi-view evidence into target-view tokens before denoising, enabling streaming scene exploration from a single image or text prompt. Practitioners working on interactive world models gain an approach that improves long-horizon consistency without explicit depth-based correspondences.
A new arXiv paper presents WorldCrafter, a video world model that learns a camera-queryable implicit 3D-aware memory to maintain consistency across long horizons and changing viewpoints. A memory encoder and pose-conditioned readout compress multi-view evidence into target-view tokens before denoising, enabling streaming scene exploration from a single image or text prompt. Practitioners working on interactive world models gain an approach that improves long-horizon consistency without explicit depth-based correspondences.
0 engagement1 sourcearxivMon, Sep 21, 2026, 05:57 PM
Developments·Paper
MoD-VLLM: Modularized Dynamic-Granularity Video LLM for Multi-Event Long Video Understanding
Researchers propose MoD-VLLM, a modular Video LLM that dynamically allocates visual token budgets across events in long videos. It addresses the tension between limited tokens and multiple key events by using adaptive capacity allocation and self-correction, improving reliability over two-stage keyframe selection methods.
Researchers propose MoD-VLLM, a modular Video LLM that dynamically allocates visual token budgets across events in long videos. It addresses the tension between limited tokens and multiple key events by using adaptive capacity allocation and self-correction, improving reliability over two-stage keyframe selection methods.
0 engagement1 sourcearxivFri, Jul 17, 2026, 09:21 AM
News·Company News
MIT and IBM launch joint computing research lab to shape future of AI
MIT and IBM announced the launch of the MIT-IBM Computing Research Lab, a collaboration aimed at advancing AI and computing research. The lab will focus on areas such as machine learning, sustainability, and startups, leveraging both institutions' expertise.
MIT and IBM announced the launch of the MIT-IBM Computing Research Lab, a collaboration aimed at advancing AI and computing research. The lab will focus on areas such as machine learning, sustainability, and startups, leveraging both institutions' expertise.
0 engagement1 sourcenews.mit.eduMon, Aug 17, 2026, 06:46 AM