{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.15404", "kind": "paper", "url": "https://arxiv.org/abs/2608.15404"}, {"id": "gh:meric-karadag/cbx-bench", "kind": "repo", "url": "https://github.com/meric-karadag/cbx-bench"}], "authors": ["Yusuf Meric Karadag", "Gulay Oklan", "Seref Baris Cagliyan", "Umut Ozdemir", "Emre Akbas"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.15404", "source_url": "https://arxiv.org/abs/2608.15404"}, "categories": ["General AI", "Vision & 3D", "cs.CV"], "description": {"en": "CBX-Bench is a benchmark for quantitatively measuring the quality of Concept Bottleneck Model (CBM) explanations. It uses a council of five open-weight multimodal LLMs to score explanations given an image and class, validated against human preferences. The benchmark maintains a leaderboard for CBM explanation quality."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.15404", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.15404", "title": "CBX-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Middle East Technical University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.15404", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.15404", "source_url": "https://openalex.org/W7203707110"}, "key": "claire-radar:2608.15404", "modality": null, "name": "CBX-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_cbx-bench_5fdbe064", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "9aefdde80e40e2fa2286f339105c26d410d8de941e233b9bb474e619c5860e6e", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2608.15404"}, "publisher": null, "released": "2026-08-15", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.15404", "source_url": "https://arxiv.org/abs/2608.15404"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-15404", "source": "claire_radar", "source_benchmark_id": "2608.15404", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Vision & 3D", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/meric-karadag/cbx-bench", "githubScope": "benchmark_repo", "githubStars": 5, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.15404", "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": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "fresh"}, "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": "effde83cc5955e30102b6fbc7361e7b38ab5ce99cfa0bc1f196823f7771a4991", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "CBX-Bench is a benchmark for quantitatively measuring the quality of Concept Bottleneck Model (CBM) explanations. It uses a council of five open-weight multimodal LLMs to score explanations given an image and class, validated against human preferences. The benchmark maintains a leaderboard for CBM explanation quality.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "Building on this validated council, we introduce CBX-Bench, a public benchmark and leaderboard: authors of new CBMs can submit their model's explanations, and CBX-Bench scores them with the council and maintains dataset-level rankings of explanation quality."}, "familyId": "bmf_825405358cac", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/meric-karadag/cbx-bench", "url": "https://github.com/meric-karadag/cbx-bench"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/meric-karadag/cbx-bench", "url": "https://huggingface.co/papers/2608.15404"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/meric-karadag/cbx-bench.", "url": "https://arxiv.org/abs/2608.15404"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/meric-karadag/cbx-bench", "url": "https://arxiv.org/abs/2608.15404"}], "scope": "benchmark_repo", "url": "https://github.com/meric-karadag/cbx-bench", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_cbx-bench_5fdbe064", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/meric-karadag/cbx-bench", "data": null, "hfPaper": "https://huggingface.co/papers/2608.15404", "paper": "https://arxiv.org/abs/2608.15404", "pdf": "https://arxiv.org/pdf/2608.15404", "project": null, "report": "https://arxiv.org/abs/2608.15404"}, "metrics": [], "motivation": "Concept Bottleneck Models (CBMs) are designed to make visual classification interpretable by expressing predictions through human-understandable concepts.", "name": "CBX-Bench", "oneLine": "CBX-Bench is a benchmark for quantitatively measuring the quality of Concept Bottleneck Model (CBM) explanations. It uses a council of five open-weight multimodal LLMs to score explanations given an image and class, validated against human preferences. The benchmark maintains a leaderboard for CBM explanation quality.", "paperTitle": "CBX-Bench: A Human-Aligned MLLM Council for Benchmarking Concept Bottleneck Model Explanations", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to the Explainable Computer Vision (eXCV) Workshop at ECCV 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2608.15404", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "Explainable Computer Vision (eXCV) Workshop at ECCV 2026", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "CBX-Bench Team", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/meric-karadag/cbx-bench"}], "ranking": {"90d": {"confidence": "Low", "coverage": 0.55, "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": 0.2391676802567412, "value": 5}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": null, "value": null}}, "confidence": "Low", "coverage": 0.55, "rank": 178, "score": 36}, "method": "attention-ranking-v13", "rank": 178, "score": 36}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-08-15", "source": {"authors": ["Yusuf Meric Karadag", "Gulay Oklan", "Seref Baris Cagliyan", "Umut Ozdemir", "Emre Akbas"], "categories": ["cs.CV"], "id": "2608.15404", "title": "CBX-Bench: A Human-Aligned MLLM Council for Benchmarking Concept Bottleneck Model Explanations", "type": "arxiv", "url": "https://arxiv.org/abs/2608.15404"}, "sourceUpdatedAt": "2026-08-15T20:21:45Z", "topics": ["cs.CV"], "venueAttempts": [{"decisionRaw": "Accepted to the Explainable Computer Vision (eXCV) Workshop at ECCV 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted to the Explainable Computer Vision (eXCV) Workshop at ECCV 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2608.15404"}], "reviewStatus": "accepted", "venueName": "Explainable Computer Vision (eXCV) Workshop at ECCV 2026"}], "whyItMatters": "CBM interpretability is often evaluated by downstream accuracy, lacking quantitative measures of explanation quality. CBX-Bench offers a human-aligned, scalable evaluation that does not require concept ground truth, enabling comparison of explanation quality across different CBMs."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
