{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2609.39533", "kind": "paper", "url": "https://arxiv.org/abs/2609.39533"}, {"id": "gh:thuais-lab/catch", "kind": "repo", "url": "https://github.com/THUAIS-Lab/CATCH"}], "authors": ["Shouli Wang", "Yanfeng Jia", "Zhihao Ou", "Zitao Su", "Ruize He", "Haotong Xie", "Hao Peng", "Juanzi Li", "Xiaozhi Wang"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2609.39533", "source_url": "https://arxiv.org/abs/2609.39533"}, "categories": ["Cybersecurity", "Language & Knowledge", "cs.CL"], "description": {"en": "During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2609.39533", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2609.39533", "title": "CATCH"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Tsinghua University", "Beihang University", "Southern University of Science and Technology", "Renmin University of China", "Shanghai University of Finance and Economics"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2609.39533", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2609.39533", "source_url": "https://openalex.org/W7215054551"}, "key": "claire-radar:2609.39533", "modality": null, "name": "CATCH", "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": "true", "origin_record_id": "bm_catch_1310e8c8", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "0f59271ca920689d1ec5b291081ff380626cd44fc9b4d146fb8a8cad9898b4d8", "relation": "introduces", "review_state": "rule-auto-admitted", "reviewed_at": "2026-10-01T07:28:29.742921Z", "source_url": "https://arxiv.org/abs/2609.39533"}, "publisher": null, "released": "2026-09-30", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2609.39533", "source_url": "https://arxiv.org/abs/2609.39533"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2609-39533", "source": "claire_radar", "source_benchmark_id": "2609.39533", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Cybersecurity"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/THUAIS-Lab/CATCH", "githubScope": "benchmark_repo", "githubStars": 2, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": 0, "hfPaperUrl": "https://huggingface.co/papers/2609.39533", "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": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "capabilities": [], "confidence": "High", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "curation": {"decisionReason": "Automatically admitted from an explicit named benchmark release with evaluation-protocol evidence at or above the configured publication-confidence threshold.", "policy": "source-evidence-v2", "reviewedAt": "2026-10-01T07:28:29.742921Z", "state": "rule-auto-admitted"}, "dataStatus": "primary-source-indexed", "demo": false, "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce CATCH, a controllable testbed for studying reward hacking in coding RL."}, "familyId": "bmf_9f925cdc4fee", "firstSeenAt": "2026-10-01", "heat": null, "id": "bm_catch_1310e8c8", "indexedAt": "2026-10-01T07:27:53.594997Z", "industrySectors": ["Cybersecurity"], "links": {"code": "https://github.com/THUAIS-Lab/CATCH", "data": null, "hfPaper": "https://huggingface.co/papers/2609.39533", "paper": "https://arxiv.org/abs/2609.39533", "pdf": "https://arxiv.org/pdf/2609.39533", "project": null, "report": "https://arxiv.org/abs/2609.39533"}, "metrics": [], "motivation": "During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking.", "name": "CATCH", "oneLine": "During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking.", "paperTitle": "CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL", "primaryDomain": "Cybersecurity", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2609.39533", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "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": 30, "historyDate": "2026-09-05", "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"normalization": "log-max", "normalized": 0.16198483906027114, "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": 218, "score": 16}, "method": "attention-ranking-v13", "rank": 218, "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": 0.1466449917496476, "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": 377, "score": 23}, "method": "attention-ranking-v13", "rank": 377, "score": 23}}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-09-30", "source": {"authors": ["Shouli Wang", "Yanfeng Jia", "Zhihao Ou", "Zitao Su", "Ruize He", "Haotong Xie", "Hao Peng", "Juanzi Li", "Xiaozhi Wang"], "categories": ["cs.CL", "cs.LG"], "id": "2609.39533", "title": "CATCH: A Controllable Analysis Testbed for Reward Hacking in Coding RL", "type": "arxiv", "url": "https://arxiv.org/abs/2609.39533"}, "sourceUpdatedAt": "2026-09-30T11:36:30Z", "topics": ["cs.CL"], "venueAttempts": []}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
