{
  "schema_version": "1.1",
  "id": "s13:https://arxiv.org/abs/2607.28627v1",
  "slug": "2607-28627v1-0pc8021",
  "url": "https://feed7.dev/p/2607-28627v1-0pc8021",
  "title": "ReToken: One Token to Improve Vision-Language Models for Visual Retrieval",
  "why_included": "ReToken adds one learned retrieval embedding to select relevant visual tokens from a cached context, improving long-image and video benchmarks while fitting training and inference on one H100.",
  "summary": "ReToken trains **one learnable embedding** as a retrieval target, selecting sparse query-relevant visual tokens from a pre-filled KV cache. It is trained on a small image-QA dataset and applied to long visual contexts where distractors otherwise degrade performance.",
  "practical_implication": "For visual agents, test retrieval over cached tokens before expanding the context processed by the model. Reported gains reach **13.4 points** for Qwen3VL-8B and **12.4 points** for InternVL3.5 on Visual Haystacks, with training and long-video inference fitting on one H100.",
  "agent_context": "ReToken trains **one learnable embedding** as a retrieval target, selecting sparse query-relevant visual tokens from a pre-filled KV cache. It is trained on a small image-QA dataset and applied to long visual contexts where distractors otherwise degrade performance.\n\nFor visual agents, test retrieval over cached tokens before expanding the context processed by the model. Reported gains reach **13.4 points** for Qwen3VL-8B and **12.4 points** for InternVL3.5 on Visual Haystacks, with training and long-video inference fitting on one H100.\n\nThe method transfers zero-shot to LVBench for an **8.0-point** gain with Qwen3VL-8B, but the abstract provides no latency, token-selection rate, or broader model coverage. The results remain benchmark evidence rather than an application-level reliability claim.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2607.28627v1",
    "published_at": "2026-07-30T17:59:56.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "context",
  "domains": [
    "image",
    "video"
  ],
  "topics": [
    "retrieval",
    "context-engineering",
    "context-caching"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The method transfers zero-shot to LVBench for an **8.0-point** gain with Qwen3VL-8B, but the abstract provides no latency, token-selection rate, or broader model coverage. The results remain benchmark evidence rather than an application-level reliability claim."
  ],
  "connected_context": {
    "meaning": "This adds a trained, token-level visual retrieval mechanism to the broader case for selecting evidence before reasoning over long context. Unlike training-free replay or external retrieval, it operates over a pre-filled visual KV cache with one learned target embedding. The benchmark gains support sparse query-relevant selection, but missing latency, selection-rate, model-breadth, and application evidence limit production conclusions.",
    "corpus_size": 297,
    "generated_at": "2026-07-31T10:07:27.765Z",
    "connections": [
      {
        "title": "ReContext: Recursive Evidence Replay as LLM Harness for Long-Context Reasoning",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.02509v1",
        "feed7_url": "https://feed7.dev/p/2607-02509v1-11vodps",
        "reason": "Both replay query-relevant evidence instead of processing all long context uniformly, but ReToken learns visual-token selection whereas ReContext is a training-free text harness."
      },
      {
        "title": "RLM: Recursive Language Models for Large Codebases - Shashi, Superagentic AI",
        "source_name": "YouTube",
        "source_url": "https://www.youtube.com/watch?v=8oyalrfwgjw",
        "feed7_url": "https://feed7.dev/p/rlm-recursive-language-models-for-large-codebases-shashi-superagentic-ai-1w21nex",
        "reason": "Both treat large context as evidence to inspect selectively and return in bounded form, extending the same design principle across visual caches and code repositories."
      },
      {
        "title": "AI on Your Lakehouse: Context Comes in Shapes, Not Queries — Zach Blumenfeld, Neo4j",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=kRkcNOsRyYg",
        "feed7_url": "https://feed7.dev/p/ai-on-your-lakehouse-context-comes-in-shapes-not-queries-zach-blumenfeld-0r4s9u5",
        "reason": "The graph-retrieval candidate reinforces that retrieval shape should match the data; ReToken supplies a specialized shape for sparse evidence inside long visual contexts."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-07-30T17:59:56.000Z",
  "modified_at": "2026-07-30T17:59:56.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2607-28627v1-0pc8021",
    "json": "https://feed7.dev/p/2607-28627v1-0pc8021.json",
    "markdown": "https://feed7.dev/p/2607-28627v1-0pc8021.md"
  }
}