{
  "schema_version": "1.0",
  "id": "archive:https://arxiv.org/abs/2607.02509v1",
  "slug": "2607-02509v1-11vodps",
  "url": "https://feed7.dev/p/2607-02509v1-11vodps",
  "title": "ReContext: Recursive Evidence Replay as LLM Harness for Long-Context Reasoning",
  "why_included": "ReContext is a training-free harness that replays query-relevant evidence from long inputs before answering, taking the best average rank across 8 long-context benchmarks up to 128K on Qwen3-4B/8B and Llama3-8B.",
  "summary": "**The gist** ReContext uses the model's own **attention-based relevance signals** to build a query-conditioned evidence pool from the full input, then replays it before generation — **training-free**, no external memory, no context pruning. It takes the **best average rank on 8 long-context datasets up to 128K** across **Qwen3-4B, Qwen3-8B, and Llama3-8B**.",
  "practical_implication": "**Why it matters** Big context windows don't mean the model uses what's in them; this is harness-level evidence that restructuring when evidence is seen beats trusting raw attention — a pattern worth stealing for retrieval and context-assembly steps in agent loops.",
  "agent_context": "**The gist** ReContext uses the model's own **attention-based relevance signals** to build a query-conditioned evidence pool from the full input, then replays it before generation — **training-free**, no external memory, no context pruning. It takes the **best average rank on 8 long-context datasets up to 128K** across **Qwen3-4B, Qwen3-8B, and Llama3-8B**.\n\n**Why it matters** Big context windows don't mean the model uses what's in them; this is harness-level evidence that restructuring when evidence is seen beats trusting raw attention — a pattern worth stealing for retrieval and context-assembly steps in agent loops.\n\n**Watch out** Gains were shown on **small open models (4B–8B)**; whether frontier models with stronger long-context behavior benefit, and what the recursive replay costs in extra tokens and latency, isn't established.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2607.02509v1",
    "published_at": null
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "context",
  "domains": [
    "research"
  ],
  "topics": [
    "context-engineering",
    "retrieval",
    "reasoning"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "Gains were shown on **small open models (4B–8B)**; whether frontier models with stronger long-context behavior benefit, and what the recursive replay costs in extra tokens and latency, isn't established."
  ],
  "lifecycle": "Current",
  "published_at": null,
  "modified_at": null,
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2607-02509v1-11vodps",
    "json": "https://feed7.dev/p/2607-02509v1-11vodps.json",
    "markdown": "https://feed7.dev/p/2607-02509v1-11vodps.md"
  }
}