{"record": {"aliases": ["ObserverBench: Testing Mechanistic Estimates for Intervention and Control"], "artifacts": [{"id": "arxiv:2609.03026", "kind": "paper", "url": "https://arxiv.org/abs/2609.03026"}, {"kind": "website", "url": "https://kwisatzh.github.io/observerbench/"}], "categories": ["General AI", "Safety", "Safety & Trustworthiness"], "description": {"en": "Benchmark framework that tests whether mechanistic interpretability observers are adequate for intervention, control, or safety tasks by reporting estimation accuracy and action loss separately."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.03026", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.03026", "title": "ObserverBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2609.03026", "modality": null, "name": "ObserverBench", "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": "true", "origin_record_id": "bm_observerbench-testing-mechanistic-estimate_c0fcde73", "origin_source": "arxiv", "recognition_confidence": "0.55", "record_sha256": "526836da4cabb75dfe1775c8a134bc7a88d3ce387cdc766e24241e529e8b5466", "relation": "evaluates_only", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-04T06:15:48.604366Z", "source_url": "https://arxiv.org/abs/2609.03026"}, "publisher": null, "released": "2026-09-02", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.03026", "source_url": "https://arxiv.org/abs/2609.03026"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-03026", "source": "claire_radar", "source_benchmark_id": "2609.03026", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["ObserverBench: Testing Mechanistic Estimates for Intervention and Control"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Safety & Trustworthiness", "attention": {"asOf": "2026-09-04", "attemptedAt": "2026-09-04T01:34:35.444253Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2609.03026", "observedAt": "2026-09-03T06:34:20.331209Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-09-04T01:34:35.444253Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Mechanistic interpretability is a high-interest area, and the benchmark's focus on action-oriented evaluation with a public project page supports moderate early attention.", "score": 55}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-04T06:15:48.604366Z", "inputHash": "9a01b91173872097aac9e45f2f263602e8c1a10bfd6deaeb0b27d56f7adbd6de", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "We present ObserverBench, a benchmark framework for testing whether an internal estimator---an observer---is adequate", "canonicalNameSource": "abstract", "decisionReason": "Formal benchmark framework with a project page and described fixed task contracts, runnable baselines, and submission format, indicating a public reuse path.", "model": "deepseek-v4-pro", "reviewedAt": "2026-09-04T06:15:48.604366Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Benchmark framework that tests whether mechanistic interpretability observers are adequate for intervention, control, or safety tasks by reporting estimation accuracy and action loss separately.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL", "named entity is a model, method, framework, or dataset used in evaluation"], "snippet": "We present ObserverBench, a benchmark framework for testing whether an internal estimator---an observer---is adequate for the intervention, control, or safety task it directs."}, "familyId": "bmf_5db486b649ec", "firstSeenAt": "2026-09-04", "heat": null, "id": "bm_observerbench-testing-mechanistic-estimate_c0fcde73", "indexedAt": "2026-09-04T00:41:56.430646Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "https://arxiv.org/abs/2609.03026", "pdf": "https://arxiv.org/pdf/2609.03026", "project": "https://kwisatzh.github.io/observerbench/", "report": "https://arxiv.org/abs/2609.03026"}, "metrics": [], "motivation": "Mechanistic interpretability is increasingly used to guide interventions such as activation steering, circuit removal, and safety monitoring.", "name": "ObserverBench", "oneLine": "Benchmark framework that tests whether mechanistic interpretability observers are adequate for intervention, control, or safety tasks by reporting estimation accuracy and action loss separately.", "paperTitle": "ObserverBench: Testing Mechanistic Estimates for Intervention and Control", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.03026", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-04T00:41:56.430646Z"}, "publications": [], "publishers": [{"name": "ObserverBench Team", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://kwisatzh.github.io/observerbench/"}], "ranking": {"30d": {"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": 30}, "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}, "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": 0.55, "relation": "evaluates_only", "releasedAt": "2026-09-02", "source": {"authors": ["Vijay Erramilli"], "categories": ["cs.LG", "cs.AI"], "id": "2609.03026", "title": "ObserverBench: Testing Mechanistic Estimates for Intervention and Control", "type": "arxiv", "url": "https://arxiv.org/abs/2609.03026"}, "sourceUpdatedAt": "2026-09-02T18:01:17Z", "topics": ["Safety"], "venueAttempts": [], "whyItMatters": "Introduces task-level evaluation for interpretability methods, showing that AUROC or estimation accuracy can misrank observers and that action loss better reflects deployment value."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
