{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.09113", "kind": "paper", "url": "https://arxiv.org/abs/2609.09113"}, {"id": "gh:trae1oung/saescientist", "kind": "repo", "url": "https://github.com/Trae1ounG/SAEScientist"}], "authors": ["Yuqiao Tan", "Shizhu He", "Jun Zhao", "Kang Liu"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.09113", "source_url": "https://arxiv.org/abs/2609.09113"}, "categories": ["Agents & Tool Use", "General AI", "cs.AI"], "description": {"en": "Evaluates AI agents performing sparse autoencoder feature interpretation on Gemma-2-9B-IT, measuring feature rank, activation separation, and steering ability across 20 tasks."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.09113", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.09113", "title": "SAEScientist-Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Chinese Academy of Sciences", "Institute of Automation", "University of Chinese Academy of Sciences"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.09113", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.09113", "source_url": "https://openalex.org/W7212060465"}, "key": "claire-radar:2609.09113", "modality": null, "name": "SAEScientist-Bench", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "High", "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_saescientist-bench_a865167b", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "c79a5b6d3455147d1677afe1a28cf8ddbc80025e69d54f1d6344f91a6ddc154a", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-09-10T00:52:28.558291Z", "source_url": "https://arxiv.org/abs/2609.09113"}, "publisher": null, "released": "2026-09-08", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.09113", "source_url": "https://arxiv.org/abs/2609.09113"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-09113", "source": "claire_radar", "source_benchmark_id": "2609.09113", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Agents & Tool Use", "attention": {"asOf": "2026-09-09", "attemptedAt": "2026-09-09T12:31:07.558041Z", "githubRepo": "https://github.com/Trae1ounG/SAEScientist", "githubScope": "benchmark_repo", "githubStars": 2, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 0, "hfPaperUrl": "https://huggingface.co/papers/2609.09113", "observedAt": "2026-09-09T12:31:07.558041Z", "signalStatus": {"githubStars": {"observedAt": "2026-09-09T12:31:07.558041Z", "state": "fresh"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-09-09T12:31:07.558041Z", "state": "fresh"}}, "status": "fresh"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The benchmark combines two niche yet trending topics, AI interpretability and agentic research, but its reliance on specific model weights and GPU requirements can restrict early adoption.", "score": 35}, "capabilities": [], "confidence": "High", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-10T00:52:28.558291Z", "inputHash": "d89bf3f3710a7494d060a9eabe77418736bfe0de3c45a611e0e62167b6da6ec6", "model": "deepseek-v4-pro"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "# SAEScientist-Bench ### Can AI Agents Conduct Autonomous SAE Interpretability Research?", "canonicalNameSource": "official_readme", "decisionReason": "Offers a defined task set, an automated scoring protocol, and public code to run the benchmark, satisfying stable scoring and reuse requirements.", "evaluationEvidence": {"reuse": "The benchmark contains 20 tasks across 17 concepts, using Gemma-2-9B-IT and Gemma Scope dictionaries with 131,072 features at layers 9 and 20. This repository provides the dataset, agent interfaces, and experiment and scoring code used to run the benchmark.", "scoring": "Scores are computed automatically during the experiment. Task scores are averaged equally within each repeat, then summarized across repeats.", "task": "Given a target concept, an agent designs contrastive probes and navigates a Gemma Scope dictionary of 131K+ features in Gemma-2-9B-IT to discover the optimal feature."}, "model": "deepseek-v4-pro", "reviewedAt": "2026-09-10T00:52:28.558291Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Evaluates AI agents performing sparse autoencoder feature interpretation on Gemma-2-9B-IT, measuring feature rank, activation separation, and steering ability across 20 tasks.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "score_submission", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "In this paper, we introduce SAEScientist-Bench to evaluate whether AI agents can act as scientists utilizing SAE tools for autonomous mechanistic discovery."}, "familyId": "bmf_b867452e1083", "firstSeenAt": "2026-09-09", "heat": null, "id": "bm_saescientist-bench_a865167b", "indexedAt": "2026-09-09T06:18:29.794093Z", "industrySectors": [], "links": {"code": "https://github.com/Trae1ounG/SAEScientist", "data": null, "paper": "https://arxiv.org/abs/2609.09113", "pdf": "https://arxiv.org/pdf/2609.09113", "project": null, "report": "https://arxiv.org/abs/2609.09113"}, "metrics": [], "motivation": "While research on recursive self-improvement (RSI) has predominantly automated model training pipelines, reliable autonomous development demands a missing pillar: post-hoc monitoring and auditing to understand what models learn and ensure safe alignment.", "name": "SAEScientist-Bench", "oneLine": "Evaluates AI agents performing sparse autoencoder feature interpretation on Gemma-2-9B-IT, measuring feature rank, activation separation, and steering ability across 20 tasks.", "paperTitle": "SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.09113", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-09-09T06:18:29.794093Z"}, "publications": [], "publishers": [{"name": "Trae1ounG", "organizationType": "community", "role": "benchmark-publisher", "sourceUrl": "https://github.com/Trae1ounG/SAEScientist"}], "ranking": {"30d": {"confidence": "High", "coverage": 0.85, "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": 0.1636259138450751, "value": 2}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.0, "value": 0}}, "confidence": "High", "coverage": 0.85, "rank": 230, "score": 16}, "method": "attention-ranking-v13", "rank": 230, "score": 16}, "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": 0.14837532069689596, "value": 2}, "hfDatasetDownloads": {"normalization": "log-max", "normalized": null, "value": null}, "hfPaperUpvotes": {"normalization": "log-max", "normalized": 0.0, "value": 0}}, "confidence": "Medium", "coverage": 0.7, "rank": 322, "score": 23}, "method": "attention-ranking-v13", "rank": 322, "score": 23}, "today": {"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": 1, "historyDate": "2026-09-08", "rank": null, "score": null, "windowDays": 1}, "level": {"components": {"githubStars": {"normalization": "percentile", "normalized": 0.8125, "percentile": 0.8125, "value": 2}, "hfDatasetDownloads": {"normalization": "percentile", "normalized": null, "percentile": null, "value": null}, "hfPaperUpvotes": {"normalization": "percentile", "normalized": 0.5, "percentile": 0.5, "value": 0}, "llmAttentionForecast": {"maxContribution": 0.25, "percentile": 0.7, "role": "bonus", "value": 70}}, "confidence": "Medium", "coverage": 0.7, "rank": 2, "score": 63}, "method": "attention-ranking-v13", "rank": 2, "score": 63}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-09-08", "source": {"authors": ["Yuqiao Tan", "Shizhu He", "Jun Zhao", "Kang Liu"], "categories": ["cs.AI", "cs.CL", "cs.LG"], "id": "2609.09113", "title": "SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?", "type": "arxiv", "url": "https://arxiv.org/abs/2609.09113"}, "sourceUpdatedAt": "2026-09-08T17:45:09Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "Provides a measurable benchmark for autonomous mechanistic interpretability, a capability needed for closed-loop AI safety research."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
