{
  "schema_version": "1.1",
  "id": "archive:https://arxiv.org/abs/2610.42116v1",
  "slug": "2610-42116v1-086agqc",
  "url": "https://feed7.dev/p/2610-42116v1-086agqc",
  "title": "ACCT: Asymmetric Cross-LLM Context Transfer",
  "why_included": "ACCT transfers selected KV cache states between differently sized models in the same family, so a target can reuse context computation. The paper reports faster prefill conversion and similar or better task performance with a smaller cache.",
  "summary": "ACCT converts a source model’s KV cache into sparse states a target model in the same family can use directly. Its Cartographer selects relevant tokens, Allocator distributes a shared KV budget, and Compiler produces the target states. The paper reports matching or better performance with **75% less KV cache**, and **6.5–21.2×** faster conversion than target prefill at about **19K tokens**.",
  "practical_implication": "For agent workflows that reuse long context across model sizes, this suggests a way to avoid repeating prefill when switching models. The approach depends on transferring cache state between models in the same family, rather than treating context as portable across arbitrary models.",
  "agent_context": "ACCT converts a source model’s KV cache into sparse states a target model in the same family can use directly. Its Cartographer selects relevant tokens, Allocator distributes a shared KV budget, and Compiler produces the target states. The paper reports matching or better performance with **75% less KV cache**, and **6.5–21.2×** faster conversion than target prefill at about **19K tokens**.\n\nFor agent workflows that reuse long context across model sizes, this suggests a way to avoid repeating prefill when switching models. The approach depends on transferring cache state between models in the same family, rather than treating context as portable across arbitrary models.\n\nThe results cover **five general-capability benchmarks** and **six context-dependent tasks**. The material does not specify implementation availability or how the method performs in production systems.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2610.42116v1",
    "published_at": "2026-10-30T17:50:49.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "context",
  "domains": [
    "research"
  ],
  "topics": [
    "context-engineering",
    "context-caching"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The results cover **five general-capability benchmarks** and **six context-dependent tasks**. The material does not specify implementation availability or how the method performs in production systems."
  ],
  "connected_context": {
    "meaning": "ACCT adds a model-side option to reuse long context across sizes within one model family: it transfers KV state directly, with reported cache and conversion savings. This complements approaches that retain or retrieve context at the harness level, but its family constraint and benchmark scope leave broader production portability unestablished.",
    "corpus_size": 974,
    "generated_at": "2026-11-02T11:02:39.748Z",
    "connections": [
      {
        "title": "Context Engineering in 2026 — Louis-François Bouchard, Omar Solano & Samridhi Vaid, Towards AI",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=WP3hjUXd918",
        "feed7_url": "https://feed7.dev/p/context-engineering-in-2026-louis-francois-bouchard-omar-solano-samridhi-1dnlyr0",
        "reason": "The full-history result favors keeping cached context when it is cheap; ACCT could reduce the cache cost of that choice when switching model sizes within a family."
      },
      {
        "title": "Better prompt caching for GPT-6",
        "source_name": "OpenAI",
        "source_url": "https://openai.com/index/better-prompt-caching-for-gpt-6",
        "feed7_url": "https://feed7.dev/p/better-prompt-caching-for-gpt-6-0mbpmnp",
        "reason": "Both target the cost of repeated context through cache reuse, but ACCT transfers KV state between same-family models while this candidate focuses on improving prompt-cache behavior."
      },
      {
        "title": "ReToken: One Token to Improve Vision-Language Models for Visual Retrieval",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.28627v1",
        "feed7_url": "https://feed7.dev/p/2607-28627v1-0pc8021",
        "reason": "Both operate on cached representations to avoid treating every context token equally; ReToken selects visual evidence for retrieval, while ACCT transfers sparse KV states across model sizes."
      },
      {
        "title": "CliffCompaction: Cost-Efficient Compaction for Long-Horizon Coding Agents",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2609.26779v1",
        "feed7_url": "https://feed7.dev/p/2609-26779v1-1nezvu9",
        "reason": "ACCT may reduce the need to compact context when moving between same-family model sizes; CliffCompaction instead bounds context by dropping or truncating source material."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-10-30T17:50:49.000Z",
  "modified_at": "2026-10-30T17:50:49.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2610-42116v1-086agqc",
    "json": "https://feed7.dev/p/2610-42116v1-086agqc.json",
    "markdown": "https://feed7.dev/p/2610-42116v1-086agqc.md"
  }
}