{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.28008", "kind": "paper", "url": "https://arxiv.org/abs/2607.28008"}], "authors": ["Yanshi Li", "Xueru Bai", "Shuman Liu", "Long Zhang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.28008", "source_url": "https://arxiv.org/abs/2607.28008"}, "categories": ["General AI", "Language & Knowledge", "cs.CL"], "description": {"en": "RepBench compiles 353 public benchmarks into 46,149 probe texts spanning 94 capabilities, with a taxonomy of 182 clusters in 13 families. It evaluates representation readout methods across 12 models under cross-benchmark transfer, providing a reusable closed-loop pipeline for capability-aligned probing."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.28008", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.28008", "title": "RepBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2607.28008", "modality": null, "name": "RepBench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "Medium", "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_repbench_7ff35873", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "f771940b09b4c41ad1ae582cd6b2baea9d2e146fd036f3071a7a4e7a26f60d13", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.28008"}, "publisher": null, "released": "2026-07-30", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.28008", "source_url": "https://arxiv.org/abs/2607.28008"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-28008", "source": "claire_radar", "source_benchmark_id": "2607.28008", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Machine Generated", "applicationDomains": ["General AI"], "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/2607.28008", "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": "Medium", "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": "d8bab187c774e445ecd97f0923d704197d7c5ce7afbffd994fc236376409ea4a", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "RepBench compiles 353 public benchmarks into 46,149 probe texts spanning 94 capabilities, with a taxonomy of 182 clusters in 13 families. It evaluates representation readout methods across 12 models under cross-benchmark transfer, providing a reusable closed-loop pipeline for capability-aligned probing.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence"], "snippet": "We present RepBench, a benchmark-grounded data layer for capability-aligned representation probing."}, "familyId": "bmf_b37ff0b1a335", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_repbench_7ff35873", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2607.28008", "paper": "https://arxiv.org/abs/2607.28008", "pdf": "https://arxiv.org/pdf/2607.28008", "project": null, "report": "https://arxiv.org/abs/2607.28008"}, "metrics": [], "motivation": "Representation engineering reads and steers capability directions in large language models, yet methods are typically evaluated on paper-specific synthetic data.", "name": "RepBench", "oneLine": "RepBench compiles 353 public benchmarks into 46,149 probe texts spanning 94 capabilities, with a taxonomy of 182 clusters in 13 families. It evaluates representation readout methods across 12 models under cross-benchmark transfer, providing a reusable closed-loop pipeline for capability-aligned probing.", "paperTitle": "RepBench: Compiling Benchmarks into Capability Representations for Large Language Models", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.28008", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "ArXiv", "organizationType": "benchmark-organization", "role": "benchmark-publisher", "sourceUrl": "https://arxiv.org/abs/2607.28008"}], "ranking": {"90d": {"confidence": "Low", "coverage": 0.0, "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": null, "value": null}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v13", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 0.85, "relation": "introduces", "releasedAt": "2026-07-30", "source": {"authors": ["Yanshi Li", "Xueru Bai", "Shuman Liu", "Long Zhang"], "categories": ["cs.CL", "cs.AI"], "id": "2607.28008", "title": "RepBench: Compiling Benchmarks into Capability Representations for Large Language Models", "type": "arxiv", "url": "https://arxiv.org/abs/2607.28008"}, "sourceUpdatedAt": "2026-08-14T06:39:38Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "RepBench addresses the lack of comparable and reproducible evaluation for representation engineering by grounding probes in multiple public benchmarks, reducing single-source bias and enabling meaningful comparison of readout methods and aggregation criteria across models."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
