{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.14541", "kind": "paper", "url": "https://arxiv.org/abs/2607.14541"}, {"id": "gh:alibaba/atrex-bench", "kind": "repo", "url": "https://github.com/alibaba/atrex-bench"}], "authors": ["Lingyun Yang", "Yuxiao Wang", "Shenghao Liang", "Linfeng Yang", "Daocheng Ying", "Chunbo You", "Rui Zhang", "Luping Wang", "Yinghao Yu", "Guodong Yang", "Liping Zhang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.14541", "source_url": "https://arxiv.org/abs/2607.14541"}, "categories": ["Agents", "Language & Knowledge", "Software & AI Compute"], "description": {"en": "Evaluates coding agents on generating GPU kernels from PyTorch references across 30 operators and 440 shapes derived from production inference traces. Scoring uses a three-stage evaluator measuring compile success, numerical correctness, and speed-of-light (SOL) efficiency against a cached roofline. The benchmark includes 4 DSL backends and supports multi-vendor GPUs."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.14541", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.14541", "title": "Atrex-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Alibaba Group (China)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.14541", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.14541", "source_url": "https://openalex.org/W7169617130"}, "key": "claire-radar:2607.14541", "modality": null, "name": "Atrex-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-10-14T09:44:51Z", "data_status": "primary-source-indexed", "display_eligible": "unknown", "origin_record_id": "bm_atrex-bench_ab05aff2", "origin_source": "arxiv", "recognition_confidence": "0.95", "record_sha256": "8a573d52266f5013cd8b51d3f99049afd9500c060f850771201d12a548644f1c", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.14541"}, "publisher": null, "released": "2026-07-16", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.14541", "source_url": "https://arxiv.org/abs/2607.14541"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-14541", "source": "claire_radar", "source_benchmark_id": "2607.14541", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["Software & AI Compute"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/alibaba/atrex-bench", "githubScope": "benchmark_repo", "githubStars": 34, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 0, "hfPaperUrl": "https://huggingface.co/papers/2607.14541", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "capabilities": [], "confidence": "High", "construction": "Original Synthetic", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "63e87ad19eac0773073fb3036c263af57bac1f75c8689e70d1acf14ca0e070ce", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates coding agents on generating GPU kernels from PyTorch references across 30 operators and 440 shapes derived from production inference traces. Scoring uses a three-stage evaluator measuring compile success, numerical correctness, and speed-of-light (SOL) efficiency against a cached roofline. The benchmark includes 4 DSL backends and supports multi-vendor GPUs.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "public artifact URL"], "snippet": "We present Atrex-Bench, a benchmark whose 30 operators and 440 shapes are sampled directly from full-cluster production inference traces of compute-limited, memory-rich GPUs."}, "familyId": "bmf_068d0e6d2ab8", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 1, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/alibaba/atrex-bench", "url": "https://github.com/alibaba/atrex-bench"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/alibaba/atrex-bench", "url": "https://huggingface.co/papers/2607.14541"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/alibaba/atrex-bench", "url": "https://arxiv.org/abs/2607.14541"}], "scope": "benchmark_repo", "url": "https://github.com/alibaba/atrex-bench", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_atrex-bench_ab05aff2", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Software & Cloud"], "links": {"code": "https://github.com/alibaba/atrex-bench", "data": null, "hfPaper": "https://huggingface.co/papers/2607.14541", "paper": "https://arxiv.org/abs/2607.14541", "pdf": "https://arxiv.org/pdf/2607.14541", "project": null, "report": "https://arxiv.org/abs/2607.14541"}, "metrics": [], "motivation": "Existing GPU kernel generation benchmarks draw problems from synthetic or curated sources that diverge from deployed workloads.", "name": "Atrex-Bench", "oneLine": "Evaluates coding agents on generating GPU kernels from PyTorch references across 30 operators and 440 shapes derived from production inference traces. Scoring uses a three-stage evaluator measuring compile success, numerical correctness, and speed-of-light (SOL) efficiency against a cached roofline. The benchmark includes 4 DSL backends and supports multi-vendor GPUs.", "paperTitle": "Are LLM-Generated GPU Kernels Production-Ready? A Trace-Driven Benchmark and Optimization Agent", "primaryDomain": "Software & AI Compute", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.14541", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "Alibaba", "organizationType": "company-research-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/alibaba/atrex-bench"}], "ranking": {"90d": {"confidence": "Medium", "coverage": 0.7, "growth": {"components": {"githubStars": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.47457505484152435, "value": 34}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.0, "value": 0}}, "confidence": "Medium", "coverage": 0.7, "rank": 130, "score": 41}, "method": "attention-ranking-v13", "rank": 130, "score": 41}}, "readiness": "Runnable", "recognitionConfidence": 0.95, "relation": "introduces", "releasedAt": "2026-07-16", "source": {"authors": ["Lingyun Yang", "Yuxiao Wang", "Shenghao Liang", "Linfeng Yang", "Daocheng Ying", "Chunbo You", "Rui Zhang", "Luping Wang", "Yinghao Yu", "Guodong Yang", "Liping Zhang"], "categories": ["cs.AI"], "id": "2607.14541", "title": "Are LLM-Generated GPU Kernels Production-Ready? A Trace-Driven Benchmark and Optimization Agent", "type": "arxiv", "url": "https://arxiv.org/abs/2607.14541"}, "sourceUpdatedAt": "2026-07-16T03:49:50Z", "topics": ["Agents"], "venueAttempts": [], "whyItMatters": "Prior GPU kernel benchmarks draw from synthetic or curated sources that diverge from deployed workloads. This benchmark provides a production-trace-driven evaluation that emphasizes operators consuming the most serving time, enabling assessment of agent performance on relevant tasks."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
