{
  "schema_version": 1,
  "builder_plan": {
    "scenario": "Your self-hosted model spends too long generating answers token by token.",
    "change": "Check whether a compatible draft model and serving integration exist before renting compute.",
    "baseline": "The same target with speculation off, then your existing drafter if you have one.",
    "steps": [
      "Locate the paper's implementation and check its license, model support, data needs, and hardware requirements. Stop here if you cannot obtain the required artifacts.",
      "Save a small set of representative product tasks and expected outcomes. Keep evaluation tasks out of any training or tuning, and choose quality and cost limits before running.",
      "Run the baseline and the proposed change on identical inputs. Keep other settings fixed and repeat paired runs to expose variation.",
      "Save per-task outputs, failures, timings, and all usage costs. Review regressions before expanding the test."
    ],
    "measure": "Accepted draft length, tokens per second, tail latency, output quality, and total GPU cost.",
    "decision": "Keep the change only if it meets your preselected quality and latency limits and improves the metric you care about after setup and operating costs. Otherwise retain the baseline. This is an exploration plan, not a validated recommendation."
  },
  "builder_review": null,
  "arxiv_id": "2609.24150",
  "title": "Acceptance-Aware Draft Model Training for Speculative Decoding",
  "summary": "The method replaces cross-entropy/KL distribution-matching objectives for training speculative decoding draft models with losses that directly optimize expected accepted token length (EAL for greedy verification, WTV for sampling), optionally followed by GRPO with simulated acceptance length as reward. It costs a draft-model training pipeline with access to target-model outputs or logits, additional implementation complexity for two decoding-mode-specific losses, and the standard overhead of running and verifying a draft model at inference. It can fail if the reader's serving stack does not support custom draft models, if the target model is API-only (preventing draft training and verification against it), or if acceptance gains do not translate to wall-clock speedup when verification overhead dominates at small batch sizes.",
  "rationale": "Acceptance-aware draft training is a targeted improvement to speculative decoding that a small team can apply to a self-hosted sub-24GB model, but its benefit depends on workload, decoding mode, and serving-stack support, so it warrants a pilot rather than immediate adoption.",
  "technique": "Acceptance-aware draft model training for speculative decoding",
  "family": "decodificacao_especulativa",
  "recommendation": "testar",
  "published": "2026-09-21",
  "as_of": "2026-10-18",
  "editorial_status": "indexed",
  "source_url": "https://arxiv.org/abs/2609.24150",
  "independent_implementations": 1,
  "total_implementations": 2,
  "stars": 878,
  "citations": 0,
  "claims": [
    {
      "claim_id": "claim-01",
      "statement": "The method replaces cross-entropy/KL distribution-matching objectives for training speculative decoding draft models with losses that directly optimize expected accepted token length (EAL for greedy verification, WTV for sampling), optionally followed by GRPO with simulated acceptance length as reward. It costs a draft-model training pipeline with access to target-model outputs or logits, additional implementation complexity for two decoding-mode-specific losses, and the standard overhead of running and verifying a draft model at inference. It can fail if the reader's serving stack does not support custom draft models, if the target model is API-only (preventing draft training and verification against it), or if acceptance gains do not translate to wall-clock speedup when verification overhead dominates at small batch sizes.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2609.24150",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "claim_id": "claim-02",
      "statement": "The paper reports that EAL and WTV losses consistently improve accepted token length over KL-based draft training across models, tasks, and decoding settings, with WTV strongest under sampling and EAL under greedy verification; no specific speedup factor is stated in the abstract.",
      "basis": "inferred",
      "result": "",
      "baseline": "",
      "conditions": "",
      "source_url": "https://arxiv.org/abs/2609.24150",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "exposure_map": [
    {
      "dimension": "quality",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compute",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "latency",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "operations",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "compatibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "security",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "data_and_training",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    },
    {
      "dimension": "reproducibility",
      "basis": "not_evaluated",
      "finding": "",
      "interpretation": "",
      "limitation": "",
      "next_check": "",
      "source_url": "",
      "source_page": null,
      "source_excerpt": ""
    }
  ],
  "risks": [],
  "minimum_test": [],
  "open_questions": [],
  "independent_tests": [],
  "lab_reproductions": [],
  "report_available": false,
  "validation_tier": "unknown",
  "evidence_tier": "unknown",
  "infrastructure_basis": "unknown",
  "equations": [],
  "equations_status": "not_fetched",
  "core_kind": "",
  "equations_fetched_at": "",
  "equations_source_url": "",
  "equations_source_sha256": "",
  "exposures_reviewed_at": "",
  "exposures_source_sha256": ""
}