{ "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": "2610.28355", "title": "Beyond the Point Mass: Acceptance-Derived Weights and Distributional Losses for Speculative Decoding", "summary": "It replaces the point-mass cross-entropy drafter loss and D-PACE's position weights with overlap-derived (TV-based) acceptance weights and a Rényi-1/2 divergence loss during drafter training. It costs only a modification to the drafter training objective, roughly a 1.5% acceptance-length gain, and no inference-time changes, extra memory, or additional models. It can fail if the gains do not transfer beyond the Qwen3-4B setting tested, if the draft and target distributions are poorly matched so the divergence signal is noisy, or if a 1.5% acceptance improvement does not justify retraining an existing drafter pipeline.", "rationale": "The improvement over D-PACE is real but small and demonstrated on a single target model, so the reader should validate it on their own workload before adopting it as the default drafter objective.", "technique": "Acceptance-weighted distributional loss for drafter training", "family": "decodificacao_especulativa", "recommendation": "testar", "published": "2026-10-20", "as_of": "2026-10-21", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.28355", "independent_implementations": 0, "total_implementations": 1, "stars": 0, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "It replaces the point-mass cross-entropy drafter loss and D-PACE's position weights with overlap-derived (TV-based) acceptance weights and a Rényi-1/2 divergence loss during drafter training. It costs only a modification to the drafter training objective, roughly a 1.5% acceptance-length gain, and no inference-time changes, extra memory, or additional models. It can fail if the gains do not transfer beyond the Qwen3-4B setting tested, if the draft and target distributions are poorly matched so the divergence signal is noisy, or if a 1.5% acceptance improvement does not justify retraining an existing drafter pipeline.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28355", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Mean acceptance length rises from 4.50 to 4.57 at T=0 and 4.20 to 4.28 at T=1 versus D-PACE on six benchmarks with a Qwen3-4B target, with consistently higher throughput.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.28355", "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": "" }