{ "schema_version": 1, "builder_plan": { "scenario": "Your coding assistant makes changes that still need substantial repair before merging.", "change": "Try the technique on held-out issues from a repository with reliable tests.", "baseline": "The same coding agent, repository snapshot, tools, and token budget.", "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": "Correct patches, regression failures, human repair time, and total cost per accepted change.", "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": "2610.03226", "title": "D2K-Bench: Can LLM Agents Turn Expert Designs into Efficient GPU Kernels?", "summary": "The method replaces prompting LLM agents with bare task descriptions by supplying structured expert guidance at three levels (algorithm, dataflow, low-level optimizations) when generating GPU kernels. It costs only the effort of writing or reusing design guidance per task, with no additional infrastructure, and is compatible with API-accessible frontier models. It can fail when guidance is task-specific and unavailable for the reader's kernels, when the agent cannot implement the described properties (the paper shows a gap between proposed and implemented design properties), and results on B200 hardware may not transfer to the reader's 24 GB GPU.", "rationale": "Attaching expert design guidance to kernel-generation prompts is a zero-infrastructure change that the paper shows substantially improves correctness and speedup across multiple models, so the reader should adopt it whenever delegating GPU kernel work to LLM agents.", "technique": "Structured expert design guidance for kernel-generating LLM agents", "family": "agentes_de_codigo", "recommendation": "adotar", "published": "2026-10-02", "as_of": "2026-10-18", "editorial_status": "indexed", "source_url": "https://arxiv.org/abs/2610.03226", "independent_implementations": 1, "total_implementations": 2, "stars": 10, "citations": null, "claims": [ { "claim_id": "claim-01", "statement": "The method replaces prompting LLM agents with bare task descriptions by supplying structured expert guidance at three levels (algorithm, dataflow, low-level optimizations) when generating GPU kernels. It costs only the effort of writing or reusing design guidance per task, with no additional infrastructure, and is compatible with API-accessible frontier models. It can fail when guidance is task-specific and unavailable for the reader's kernels, when the agent cannot implement the described properties (the paper shows a gap between proposed and implemented design properties), and results on B200 hardware may not transfer to the reader's 24 GB GPU.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.03226", "source_page": null, "source_excerpt": "" }, { "claim_id": "claim-02", "statement": "Providing expert design guidance (L1 algorithmic insights, L2 dataflow design, L3 low-level optimizations) raises correctness from 93.1% to 98.5% across 130 model-task pairs and increases geometric mean speedup from 1.69x to 2.49x for the three frontier models, a 1.47x improvement in achieved kernel speedup.", "basis": "inferred", "result": "", "baseline": "", "conditions": "", "source_url": "https://arxiv.org/abs/2610.03226", "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": "" }