{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.21023", "kind": "paper", "url": "https://arxiv.org/abs/2606.21023"}], "authors": ["Zhenting Zhu", "Lucas Thai", "Shan Yu", "Yicheng Liu", "Yifan Qiao", "Chenxi Wang", "Harry Xu", "Junyi Shu"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.21023", "source_url": "https://arxiv.org/abs/2606.21023"}, "categories": ["Finance", "Language & Knowledge", "cs.LG"], "description": {"en": "MCR-Bench is a benchmark for evaluating reproducibility in mission-critical LLM tasks, measuring output consistency across heterogeneous hardware."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.21023", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.21023", "title": "MCR-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of California, Berkeley", "University of Chinese Academy of Sciences"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.21023", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.21023", "source_url": "https://openalex.org/W7165676502"}, "key": "claire-radar:2606.21023", "modality": null, "name": "MCR-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_mcr-bench_10494b73", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "e1eb2c73f5438854c6041180d82fa1e7fc709bf9cc0b3836eaa2248aabd9480b", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.21023"}, "publisher": null, "released": "2026-06-19", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.21023", "source_url": "https://arxiv.org/abs/2606.21023"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-21023", "source": "claire_radar", "source_benchmark_id": "2606.21023", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Finance"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2606.21023", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "d0e1eb36a5bf2c868962434555e77c8adc1ca4aef192bf057ce81e714bd3b6a9", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "MCR-Bench is a benchmark for evaluating reproducibility in mission-critical LLM tasks, measuring output consistency across heterogeneous hardware.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact named benchmark artifact released in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence"], "snippet": "To evaluate our approach practically, we introduce MCR-Bench, a benchmark targeting reproducibility in mission-critical tasks."}, "familyId": "bmf_0ade0c7988a9", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_mcr-bench_10494b73", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": ["Financial Services"], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2606.21023", "paper": "https://arxiv.org/abs/2606.21023", "pdf": "https://arxiv.org/pdf/2606.21023", "project": null, "report": "https://arxiv.org/abs/2606.21023"}, "metrics": [], "motivation": "As Large Language Models (LLMs) deploy into mission-critical domains (e.g., finance, medicine, and law), output reproducibility has become a strict system requirement.", "name": "MCR-Bench", "oneLine": "MCR-Bench is a benchmark for evaluating reproducibility in mission-critical LLM tasks, measuring output consistency across heterogeneous hardware.", "paperTitle": "Demystifying Numerical Instability in LLM Inference: Achieving Reproducible Inference for Mission-Critical Tasks with HEAL", "primaryDomain": "Finance", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.21023", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {}, "readiness": "Paper only", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-19", "source": {"authors": ["Zhenting Zhu", "Lucas Thai", "Shan Yu", "Yicheng Liu", "Yifan Qiao", "Chenxi Wang", "Harry Xu", "Junyi Shu"], "categories": ["cs.LG"], "id": "2606.21023", "title": "Demystifying Numerical Instability in LLM Inference: Achieving Reproducible Inference for Mission-Critical Tasks with HEAL", "type": "arxiv", "url": "https://arxiv.org/abs/2606.21023"}, "sourceUpdatedAt": "2026-06-19T01:21:10Z", "topics": ["cs.LG"], "venueAttempts": [], "whyItMatters": "LLM deployments in finance, medicine, and law require reproducible outputs. MCR-Bench addresses the lack of standardized evaluation for numerical instability in 16-bit inference, helping practitioners select methods that balance reproducibility and performance."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
