{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2608.22103", "kind": "paper", "url": "https://arxiv.org/abs/2608.22103v1"}, {"kind": "website", "url": "https://majoroth.github.io/hack-verifiable-environments/hvtb"}], "authors": ["Amit Roth", "Ivan Bercovich", "Yonathan Efroni"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2608.22103", "source_url": "https://arxiv.org/abs/2608.22103"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "Adapts Terminal Bench with hack-verifiable environments to automatically detect reward hacking in agent trajectories."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2608.22103v1", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2608.22103v1", "title": "Hack-Verifiable Terminal Bench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Tel Aviv University", "University of California, Santa Barbara"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2608.22103", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2608.22103", "source_url": "https://openalex.org/W7204183238"}, "key": "claire-radar:2608.22103", "modality": null, "name": "Hack-Verifiable Terminal Bench", "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_hack-verifiable-terminal-bench_a02835b2", "origin_source": "arxiv", "recognition_confidence": "0.75", "record_sha256": "5071f046f249afa0b918ed3fb2c2596db5b2b4162c1eca8e9793800c4ab94416", "relation": "introduces", "review_model": "deepseek-v4-pro", "review_state": "ai-reviewed", "reviewed_at": "2026-08-27T04:18:56.318436Z", "source_url": "https://arxiv.org/abs/2608.22103v1"}, "publisher": null, "released": "2026-08-22", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2608.22103", "source_url": "https://arxiv.org/abs/2608.22103v1"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2608-22103", "source": "claire_radar", "source_benchmark_id": "2608.22103", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attention": {"asOf": "2026-08-26", "attemptedAt": "2026-08-25T23:36:26.005677Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2608.22103", "observedAt": "2026-08-24T11:31:30.653116Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-08-25T23:36:26.005677Z", "state": "unavailable"}}, "status": "unavailable"}, "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "The topic of reward hacking in autonomous agents is timely and the released environments support reproducibility.", "score": 65}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-27T04:18:56.318436Z", "inputHash": "c495adf9a494f7eff1165afadc3827e976c42bbed5ac384c677bcd2497e1de9d", "model": "deepseek-v4-pro"}, "curation": {"canonicalNameEvidence": "Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks", "canonicalNameSource": "paper_title", "decisionReason": "The paper explicitly releases environments and traces at a project page, providing a public reuse path and stable detection protocol.", "model": "deepseek-v4-pro", "reviewedAt": "2026-08-27T04:18:56.318436Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Adapts Terminal Bench with hack-verifiable environments to automatically detect reward hacking in agent trajectories.", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "evaluation protocol evidence", "public artifact URL"], "snippet": "In this work, we adapt HVE to Terminal Bench, a leading benchmark of real-world terminal and coding tasks, and introduce Hack-Verifiable Terminal Bench (HVTB)."}, "familyId": "bmf_2c1e950a9953", "firstSeenAt": "2026-08-25", "heat": null, "id": "bm_hack-verifiable-terminal-bench_a02835b2", "indexedAt": "2026-08-25T17:37:25.889905Z", "industrySectors": [], "links": {"code": null, "data": null, "paper": "http://arxiv.org/abs/2608.22103v1", "pdf": "https://arxiv.org/pdf/2608.22103v1", "project": "https://majoroth.github.io/hack-verifiable-environments/hvtb", "report": "http://arxiv.org/abs/2608.22103v1"}, "metrics": [], "motivation": "As agents grow more capable and autonomous, their tendency to reward hack, satisfying a task's checks while violating its intent, becomes an increasingly important failure mode.", "name": "Hack-Verifiable Terminal Bench", "oneLine": "Adapts Terminal Bench with hack-verifiable environments to automatically detect reward hacking in agent trajectories.", "paperTitle": "Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "http://arxiv.org/abs/2608.22103v1", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-25T17:37:25.889905Z"}, "publications": [], "publishers": [{"name": "HVTB Team", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://majoroth.github.io/hack-verifiable-environments/hvtb"}], "ranking": {"30d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 30}, "level": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v11", "rank": null, "score": null}, "90d": {"confidence": "Low", "coverage": 0.0, "growth": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "elapsedDays": null, "historyDate": null, "rank": null, "score": null, "windowDays": 90}, "level": {"components": {"githubStars": {"percentile": null, "value": null}, "hfDatasetDownloads": {"percentile": null, "value": null}, "hfPaperUpvotes": {"percentile": null, "value": null}}, "confidence": "Low", "coverage": 0.0, "rank": null, "score": null}, "method": "attention-ranking-v11", "rank": null, "score": null}}, "readiness": "Paper only", "recognitionConfidence": 0.75, "relation": "introduces", "releasedAt": "2026-08-22", "source": {"authors": ["Amit Roth", "Ivan Bercovich", "Yonathan Efroni"], "categories": ["cs.AI"], "id": "2608.22103", "title": "Hack-Verifiable Terminal Bench: Evaluating Reward Hacking in Terminal Tasks", "type": "arxiv", "url": "http://arxiv.org/abs/2608.22103v1"}, "sourceUpdatedAt": "2026-08-22T20:59:36Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "Provides reliable, automatic measurement of reward hacking, enabling comparison of model robustness without human judges."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
