{ "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": "2608.26004", "title": "AsymSpec: Context-Asymmetric Speculative Decoding for Agentic LLMs", "summary": "AsymSpec replaces the standard lossy choice between input compression and full-context inference by letting a lightweight drafter read the full input while the large verifier generates against a compressed view, fusing their logits with a contrastive delta-fusion and a divergence-aware acceptance gate. The cost is running two models (drafter plus verifier) with dual context representations and a non-standard decoding loop, which increases memory footprint and implementation complexity relative to plain compression or off-the-shelf speculative decoding served through third-party APIs. It can fail when the divergence gate misjudges draft reliability, collapsing acceptance rates and erasing the speedup, and the reported gains are on isolated capabilities and two benchmarks, so end-to-end agentic workloads with heavy tool-call branching may see smaller benefits.", "rationale": "The approach is implementable on a single 24 GB GPU with a small drafter plus verifier and directly targets the compression-versus-accuracy trade-off common in agentic pipelines, but its custom fusion and gating logic and its unvalidated behavior on the reader's specific workloads mean it warrants a controlled pilot rather than immediate production adoption.", "technique": "Context-asymmetric speculative decoding", "family": "decodificacao_especulativa", "recommendation": "testar", "published": "2026-08-26", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2608.26004", "independent_implementations": 2, "total_implementations": 3, "stars": 1, "citations": 0, "claims": [ { "claim_id": "claim-01", "statement": "AsymSpec replaces the standard lossy choice between input compression and full-context inference by letting a lightweight drafter read the full input while the large verifier generates against a compressed view, fusing their logits with a contrastive delta-fusion and a divergence-aware acceptance gate. The cost is running two models (drafter plus verifier) with dual context representations and a non-standard decoding loop, which increases memory footprint and implementation complexity relative to plain compression or off-the-shelf speculative decoding served through third-party APIs. It can fail when the divergence gate misjudges draft reliability, collapsing acceptance rates and erasing the speedup, and the reported gains are on isolated capabilities and two benchmarks, so end-to-end agentic workloads with heavy tool-call branching may see smaller benefits.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2608.26004", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Reports 1.3-1.7x throughput speedups at 0.2-0.3x compute cost on isolated text capabilities, while retaining approximately 90% of full-context accuracy on average.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2608.26004", "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": [ { "anchor": "S3.E2", "label": "(2)", "section": "3 Method", "role": "proposed_method", "latex": "d^{\\prime}_{i}\\;=\\;\\arg\\max\\bigl(t_{i}+\\beta\\,\\delta_{i}\\bigr),\\vskip-5.0pt", "mathml": "di′=arg⁡max⁡(ti+β​δi),", "context": "where ai,bi∈ℝ|𝒱| are the per-position logits. Subtracting bi from ai removes the drafter’s context-independent preferences, isolating the shift induced by the additional context. Upon draft rejection, δi is fused into the verifier’s distribution:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E3", "label": "(3)", "section": "3 Method", "role": "proposed_method", "latex": "\\begin{gathered}\\gamma_{\\text{eff}}(i)\\;=\\;\\gamma\\cdot\\exp(-D_{i}),\\\\\nD_{i}\\;=\\;\\mathrm{JSD}\\bigl(\\mathrm{softmax}(a_{i})\\,\\|\\,\\mathrm{softmax}(b_{i})\\bigr).\\end{gathered}\\vskip-5.0pt", "mathml": "γeff​(i)=γ⋅exp⁡(−Di),Di=JSD(softmax(ai)∥softmax(bi)).", "context": "When the full and compressed views diverge sharply, a fixed acceptance threshold γ usually over-rejects the drafts and wastes context-gain signals. We replace it with a threshold that relaxes proportionally to context divergence, instantiating the per-position divergence Di as the Jensen–Shannon divergence (JSD):", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" }, { "anchor": "S3.E4", "label": "(4)", "section": "3 Method", "role": "proposed_method", "latex": "[\\mathrm{softmax}(t_{i})]_{d_{i}}\\;>\\;\\gamma_{\\text{eff}}(i)\\cdot[\\mathrm{softmax}(b_{i})]_{d_{i}}.\\vskip-5.0pt", "mathml": "[softmax⁡(ti)]di>γeff​(i)⋅[softmax⁡(bi)]di.", "context": "We choose JSD specifically for its strict upper bound, which guarantees γeff∈[γ/2,γ] without clipping or introducing additional hyperparameters; the exponential form is the unique solution to a multiplicative-composition axiom (derivation in Appendix A). A large Di signals context-induced divergence (not capacity-induced)—these are the positions where the compressed verifier most likely under-scores useful full-context drafts, so relaxing acceptance there is appropriate. The drafted token di is accepted iff:", "headline": "", "explanation": "", "objective": "", "symbols": {}, "evidence": "", "evidence_url": "", "source_url": "" } ], "equations_status": "selected", "core_kind": "formula", "equations_fetched_at": "2026-09-21", "equations_source_url": "", "equations_source_sha256": "", "exposures_reviewed_at": "", "exposures_source_sha256": "" }