{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-19-016",
    "source_story_id": "tmp-brief-dlite-history-free-decoding",
    "edition_id": "mp-2026-10-19-morning-0101",
    "edition_url": "https://themachinepress.com/edition/2026-10-19",
    "position": 18,
    "story_type": "ticker",
    "section": "infrastructure",
    "editorial_classification": "editorial",
    "headline": "The Draft Model Stops Carrying the Whole Conversation",
    "slug": "the-draft-model-stops-carrying-the-whole-conversation",
    "dek": "DLite reuses target-model hidden states for local speculative prediction instead of maintaining its own growing historical cache.",
    "summary": "DLite reuses target-model hidden states for local speculative prediction instead of maintaining its own growing historical cache.",
    "body_text": "Long agent sessions can erode speculative-decoding acceptance while making a draft model's cache expensive. DLite performs local prediction from target-model hidden states without explicitly preserving the full history. On the authors' general, long-context and agentic tests, it produced acceptance lengths of two to six and outperformed the compared DSpark method by as much as 70 percent. The gains depend on the tested models and serving stack.",
    "why_it_matters": "DLite reuses target-model hidden states for local speculative prediction instead of maintaining its own growing historical cache.",
    "limitations": [],
    "importance": 7,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-19-016/the-draft-model-stops-carrying-the-whole-conversation",
    "json_url": "https://themachinepress.com/story/mp-2026-10-19-016.json",
    "first_published_at": "2026-10-19T09:00:00.000-04:00",
    "modified_at": "2026-10-19T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-19-016-001",
        "text": "DLite reuses target-model hidden states for local speculative prediction instead of maintaining its own growing historical cache.",
        "source_ids": [
          "source-2026-10-19-018"
        ],
        "qualification": null
      }
    ],
    "source_ids": [
      "source-2026-10-19-018"
    ],
    "tags": [
      "speculative decoding",
      "agentic AI",
      "inference"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-19-018",
      "title": "arXiv preprint 2610.23787, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.23787v1",
      "canonical_url": "https://arxiv.org/abs/2610.23787v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-16T13:18:54.000-04:00",
      "accessed_at": "2026-10-19T08:27:48.782-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-19-016-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Draft Model Stops Carrying the Whole Conversation",
    "publisher": "The Machine Press",
    "published_at": "2026-10-19T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-19-016/the-draft-model-stops-carrying-the-whole-conversation"
  }
}