{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.14285", "kind": "paper", "url": "https://arxiv.org/abs/2607.14285"}], "authors": ["Aryan Keluskar", "Amrita Bhattacharjee", "Huan Liu"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.14285", "source_url": "https://arxiv.org/abs/2607.14285"}, "categories": ["Code & Software", "General AI", "Safety"], "description": {"en": "The evaluation object is a set of 128 scenarios across 16 domains for tool-calling LLM agents in regulated industries, assessing conflicts between safety-aligned values and deployment instructions. The task involves processing confidential documents and measuring override behavior."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.14285", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.14285", "title": "ToolAlignBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Arizona State University"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.14285", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.14285", "source_url": "https://openalex.org/W7169609852"}, "key": "claire-radar:2607.14285", "modality": null, "name": "ToolAlignBench", "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": "unknown", "origin_record_id": "bm_toolalignbench_7a0c33d1", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "9107730e8a80381d4d3ea1547028a32591761dfe2515adae601e2cc7f0fd9c63", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.14285"}, "publisher": null, "released": "2026-07-15", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.14285", "source_url": "https://arxiv.org/abs/2607.14285"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-14285", "source": "claire_radar", "source_benchmark_id": "2607.14285", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Code & Software", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": null, "githubScope": null, "githubStars": null, "hfDailySubmittedAt": null, "hfDataset": null, "hfDatasetDownloads": null, "hfDatasetLikes": null, "hfPaperUpvotes": null, "hfPaperUrl": "https://huggingface.co/papers/2607.14285", "observedAt": "2026-10-04T03:02:42.754940Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "state": "unavailable"}}, "status": "unavailable"}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "6bad161e55b7b44e9985b91f2d6762d5b3e449566835eeffc74e38c7ee1bbe6d", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "The evaluation object is a set of 128 scenarios across 16 domains for tool-calling LLM agents in regulated industries, assessing conflicts between safety-aligned values and deployment instructions. The task involves processing confidential documents and measuring override behavior.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["coined title prefix ending in Bench or Benchmark", "benchmark release stated in one sentence"], "snippet": "To empirically verify this phenomenon, we build a benchmark of 128 scenarios across 16 domains."}, "familyId": "bmf_95022812d16e", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_toolalignbench_7a0c33d1", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2607.14285", "paper": "https://arxiv.org/abs/2607.14285", "pdf": "https://arxiv.org/pdf/2607.14285", "project": null, "report": "https://arxiv.org/abs/2607.14285"}, "metrics": [], "motivation": "Safety alignment in LLMs aims to align models with human values, but which values take precedence when they conflict?", "name": "ToolAlignBench", "oneLine": "The evaluation object is a set of 128 scenarios across 16 domains for tool-calling LLM agents in regulated industries, assessing conflicts between safety-aligned values and deployment instructions. The task involves processing confidential documents and measuring override behavior.", "paperTitle": "ToolAlignBench: Investigating Alignment Conflicts in Tool-Calling Enabled LLMs", "primaryDomain": "General AI", "publication": {"evidence": "Accepted to the Pluralistic Alignment Workshop at ICML 2026", "evidenceLevel": "author-claim", "evidenceUrl": "https://arxiv.org/abs/2607.14285", "source": "arxiv-comments", "status": "acceptance_claimed", "venue": "Pluralistic Alignment Workshop at ICML 2026", "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {"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": 0.85, "relation": "introduces", "releasedAt": "2026-07-15", "source": {"authors": ["Aryan Keluskar", "Amrita Bhattacharjee", "Huan Liu"], "categories": ["cs.SE", "cs.AI"], "id": "2607.14285", "title": "ToolAlignBench: Investigating Alignment Conflicts in Tool-Calling Enabled LLMs", "type": "arxiv", "url": "https://arxiv.org/abs/2607.14285"}, "sourceUpdatedAt": "2026-07-15T18:48:49Z", "topics": ["Safety"], "venueAttempts": [{"decisionRaw": "Accepted to the Pluralistic Alignment Workshop at ICML 2026", "evidence": [{"level": "author-claim", "observedAt": "2026-10-14T01:46:58.058199Z", "rawValue": "Accepted to the Pluralistic Alignment Workshop at ICML 2026", "sourceType": "arxiv-comments", "sourceUrl": "https://arxiv.org/abs/2607.14285"}], "reviewStatus": "accepted", "venueName": "Pluralistic Alignment Workshop at ICML 2026"}], "whyItMatters": "The evaluation gap is the lack of tests for conflicting value systems in agentic tool use. Practical value lies in identifying liability risks and tuning alignment strategies for regulated deployments."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
