Newshot
Papers Still developing Event #01M7J58JRRJVBF16CCN2NZP94G

Token2Tile: Token-Wise Sparse Attention with Tile Execution through a Low-Rank Space for Video DiTs

Event overview(model summary, not the source's own text)

arXiv 于 2611 期发布了三篇关于稀疏注意力机制的研究预印本。第一篇提出 Token2Tile,一种面向视频扩散 Transformer 的按 token 稀疏注意力方法,通过低秩空间中的 tile 执行来降低注意力二次成本。第二篇题为《Lost in a Few Heads》,诊断并修复稀疏注意力中的块选择问题,指出稀疏注意力模型在固定预算下为每个查询读取少量键块,可能损失大量长上下文检索能力。第三篇提出 RISA,一种面向块扩散语言模型的可恢复稀疏注意力。三篇均为 arXiv 新公告,尚未见到同行评审或后续实验结论。

This summary was distilled by a model from the 3 reports below; it is not the source text. Treat the reports themselves as the facts.

How this story unfolded

Left: what it concluded — the key points the model distilled from the reports below, each one listing the reports it rests on. Right: what happened — ordered by time, each with a short summary so you can follow the thread without clicking through. The numbered links work both ways: tap a source under a point to jump to that report, tap the cited-by chip on a report to jump back to the point.

Key points4 items Model summary · ordered by when the evidence was published

  1. Point 01 11/16, 13:00

    Token2Tile 针对视频扩散 Transformer 的注意力二次成本问题,提出按 token 稀疏化并在低秩空间做 tile 执行。

  2. Point 02 11/16, 13:00

    三篇论文均为 arXiv 2611 期新公告(Announce Type: new)。

  3. Point 03 11/18, 13:00

    《Lost in a Few Heads》指出稀疏注意力在固定预算下仅读取少量键块,可能损失长上下文检索能力,并提出诊断与修复方法。

  4. Point 04 11/18, 13:00

    RISA 面向块扩散语言模型,提出可恢复的稀疏注意力机制。

ReportsBy publish time · newer further down

  1. November 16, 2026 · Mon
  2. Start

    Token2Tile: Token-Wise Sparse Attention with Tile Execution through a Low-Rank Space for Video DiTs

    Original summary arXiv:2611.17930v1 Announce Type: new Abstract: Attention is essential for modeling spatio-temporal dependencies in video diffusion transformers, but its quadratic cost makes it a

    arXiv cs.CV Original ↗
    Cited by Point 01 Point 02 Cited
  3. November 18, 2026 · Wed
  4. +2d 0h

    RISA: Recoverable Sparse Attention for Block-Diffusion Language Models

    Original summary arXiv:2611.22191v1 Announce Type: new Abstract: Block-wise diffusion language models (DLMs) refine multiple tokens in parallel, offering an alternative to token-by-token autoregres

    arXiv cs.CL Original ↗
    Cited by Point 02 Point 04 Cited
  5. +2d 0h

    Lost in a Few Heads: Diagnosing and Repairing Block Selection in Sparse Attention

    Original summary arXiv:2611.21917v1 Announce Type: new Abstract: Sparse attention models read a few blocks of keys for every query at a fixed budget and can lose much of their long-context retrieva

    arXiv cs.LG Original ↗
    Cited by Point 02 Point 03 Cited
  6. Still developing · later reports will keep extending this axis