{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "lange-et-al-robust-kernel-discovery-2025",
    "title": "Lange and colleagues with LLMs: robustly evaluated GPU-kernel search",
    "summary": "Combined language-model translation, evolutionary kernel search, model-based pre-screening and GPU profiling with a benchmark varying initialization and input settings, reporting optimized forward and backward kernels while exposing inflated gains from benchmark loopholes. Reported result; not independently verified by WRS. Ranked using a provisional grade and reference estimate for the exact claimed result.",
    "boundary": "Model-based soft verification is a fallible filter, not formal proof; hardware comparisons still test finite inputs and tolerances. Speedups depend on baseline, shape and GPU, and backward kernels remain harder. Removing contaminated tasks substantially reduces headline gains.",
    "area": "AI-assisted GPU programming",
    "topic": "Computing & AI",
    "year": 2025,
    "lane": "AI",
    "result": "ALGORITHM",
    "package_id": "lange-et-al-robust-kernel-discovery-2025",
    "target": {
      "record_id": "urn:uuid:085ddbc7-17d4-427b-afc9-e3df5e1fcfc7",
      "version": 1,
      "content_hash": "sha256:8fa710360bf02d3d780c098c022d8bbc86447d631dd8545cf9072fd176630128"
    },
    "humans": [
      "Robert Tjarko Lange",
      "Qi Sun",
      "Aaditya Prasad",
      "Maxence Faldor",
      "Yujin Tang",
      "David Ha"
    ],
    "authors_complete": true,
    "human_role_note": "Named source authors and agent execution are recorded separately; individual effort shares are unknown.",
    "agents": [
      {
        "provider": "OpenAI",
        "product": "GPT and reasoning-model ensemble",
        "model": "GPT-4.1; o3; o4-mini",
        "role": "Source-reported kernel proposals and screening within the human-designed search; exact API snapshots not specified",
        "assurance": "SOURCE_REPORTED",
        "source": "cur20261017-arxiv-2509-14279v1"
      },
      {
        "provider": "Anthropic",
        "product": "Claude",
        "model": "Claude Sonnet 3.7",
        "role": "Source-reported member of proposal/verification ensembles; individual kernel credit not isolated",
        "assurance": "SOURCE_REPORTED",
        "source": "cur20261017-arxiv-2509-14279v1"
      },
      {
        "provider": "Google DeepMind",
        "product": "Gemini",
        "model": "Gemini 2.5 Pro",
        "role": "Source-reported member of model-ensemble ablation; not attributed to every kernel",
        "assurance": "SOURCE_REPORTED",
        "source": "cur20261017-arxiv-2509-14279v1"
      }
    ],
    "related": [],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "cur20261017-arxiv-2509-14279v1",
        "title": "Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimization",
        "url": "https://arxiv.org/abs/2509.14279v1",
        "date": "2025-09-16",
        "kind": "PRIMARY_RESEARCH_PAPER",
        "access": "WEB_READ",
        "checked_on": "2026-10-17"
      }
    ],
    "geography": null,
    "points": null,
    "published_grade": null,
    "assessment_disposition": {
      "status": "CLAIM_UNVERIFIED",
      "label": "Reported result · unverified",
      "reason": "Provisional grade and reference score included in rankings; scientific correctness has not been independently established.",
      "target": {
        "record_id": "urn:uuid:085ddbc7-17d4-427b-afc9-e3df5e1fcfc7",
        "version": 1,
        "content_hash": "sha256:8fa710360bf02d3d780c098c022d8bbc86447d631dd8545cf9072fd176630128"
      },
      "edition": "ratings-e039106f30dec177"
    },
    "conditional_rating": {
      "id": "lange-et-al-robust-kernel-discovery-2025",
      "target": {
        "record_id": "urn:uuid:085ddbc7-17d4-427b-afc9-e3df5e1fcfc7",
        "version": 1,
        "content_hash": "sha256:8fa710360bf02d3d780c098c022d8bbc86447d631dd8545cf9072fd176630128"
      },
      "band": "B",
      "modifier": "+",
      "points": "35000",
      "range": {
        "lower": "15000",
        "upper": "80000"
      },
      "status": "CONDITIONAL_IF_VERIFIED",
      "counts_as_achieved": false,
      "scientific_correctness_certified": false,
      "scope": "Model-based soft verification is a fallible filter, not formal proof; hardware comparisons still test finite inputs and tolerances. Speedups depend on baseline, shape and GPU, and backward kernels remain harder. Removing contaminated tasks substantially reduces headline gains."
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-22T00:00:00Z",
    "publisher": {
      "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33",
      "role": "WRS_EDITORIAL_PUBLICATION"
    },
    "record": {
      "contributors": [],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimization",
            "url": "https://arxiv.org/abs/2509.14279v1"
          }
        }
      ],
      "origin": "AGENT_NATIVE",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "Model-based soft verification is a fallible filter, not formal proof; hardware comparisons still test finite inputs and tolerances. Speedups depend on baseline, shape and GPU, and backward kernels remain harder. Removing contaminated tasks substantially reduces headline gains.",
        "domain": "AI-assisted GPU programming"
      },
      "sources": [
        {
          "citation": "Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimization",
          "url": "https://arxiv.org/abs/2509.14279v1"
        }
      ],
      "statement": "# Lange and colleagues with LLMs: robustly evaluated GPU-kernel search\n\nCombined language-model translation, evolutionary kernel search, model-based pre-screening and GPU profiling with a benchmark varying initialization and input settings, reporting optimized forward and backward kernels while exposing inflated gains from benchmark loopholes. Reported result; not independently verified by WRS. Ranked using a provisional grade and reference estimate for the exact claimed result.\n\n## Scientific scope\n\nModel-based soft verification is a fallible filter, not formal proof; hardware comparisons still test finite inputs and tolerances. Speedups depend on baseline, shape and GPU, and backward kernels remain harder. Removing contaminated tasks substantially reduces headline gains.\n\n## Reading status\n\nThis is a source-linked WRS catalogue note. A full explanatory guide and independent scientific reconstruction have not been completed for this entry. Read the identified original work for its complete narrative and arguments.\n\n## People and systems\n\nRobert Tjarko Lange; Qi Sun; Aaditya Prasad; Maxence Faldor; Yujin Tang; David Ha\n\nNamed source authors and agent execution are recorded separately; individual effort shares are unknown.\n\nGPT and reasoning-model ensemble / OpenAI / GPT-4.1; o3; o4-mini: Source-reported kernel proposals and screening within the human-designed search; exact API snapshots not specified (source reported).\n\nClaude / Anthropic / Claude Sonnet 3.7: Source-reported member of proposal/verification ensembles; individual kernel credit not isolated (source reported).\n\nGemini / Google DeepMind / Gemini 2.5 Pro: Source-reported member of model-ensemble ablation; not attributed to every kernel (source reported).\n\n## Original sources\n\n- [Towards Robust Agentic CUDA Kernel Benchmarking, Verification, and Optimization](https://arxiv.org/abs/2509.14279v1)\n",
      "summary": "Combined language-model translation, evolutionary kernel search, model-based pre-screening and GPU profiling with a benchmark varying initialization and input settings, reporting optimized forward and backward kernels while exposing inflated gains from benchmark loopholes. Reported result; not independently verified by WRS. Ranked using a provisional grade and reference estimate for the exact claimed result.",
      "title": "Lange and colleagues with LLMs: robustly evaluated GPU-kernel search"
    },
    "record_id": "urn:uuid:085ddbc7-17d4-427b-afc9-e3df5e1fcfc7",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:085ddbc7-17d4-427b-afc9-e3df5e1fcfc7",
      "version": 1,
      "content_hash": "sha256:8fa710360bf02d3d780c098c022d8bbc86447d631dd8545cf9072fd176630128"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:9d28977505918eb73ec672263480f3feb6aece8a0161987c310130b3c0c13d34"
  },
  "target": {
    "record_id": "urn:uuid:085ddbc7-17d4-427b-afc9-e3df5e1fcfc7",
    "version": 1,
    "content_hash": "sha256:8fa710360bf02d3d780c098c022d8bbc86447d631dd8545cf9072fd176630128"
  },
  "published_grade": null,
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": null,
  "reading_status": "SOURCE_LINKED_NOTE"
}