{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2607.10059", "kind": "paper", "url": "https://arxiv.org/abs/2607.10059"}], "authors": ["Xun Liu", "Yi Evie Zhang", "Vira Kasprova", "Parisa Rabbani", "Pardis Sadat Zahraei", "Tianyu Zhang", "Ali Ebrahimpour-Boroojeny", "Varun Chandrasekaran"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2607.10059", "source_url": "https://arxiv.org/abs/2607.10059"}, "categories": ["Agents", "General AI", "Language & Knowledge", "Reasoning"], "description": {"en": "AgentAbstain is a paired-task benchmark for evaluating LLM agents' ability to abstain from acting in scenarios such as ambiguity, conflicting constraints, or tool failures. It includes 263 paired tasks across 42 sandbox environments, with a proposed pipeline for generating fresh task instances."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2607.10059", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2607.10059", "title": "AgentAbstain"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["University of Illinois Urbana-Champaign"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2607.10059", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2607.10059", "source_url": "https://openalex.org/W7168252655"}, "key": "claire-radar:2607.10059", "modality": null, "name": "AgentAbstain", "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_agentabstain_d94f5610", "origin_source": "arxiv", "recognition_confidence": "0.85", "record_sha256": "89c05f129f1c35cbf1095f19d8b688ff2dd7b89c6e612764760904d804121fd5", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2607.10059"}, "publisher": null, "released": "2026-07-11", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2607.10059", "source_url": "https://arxiv.org/abs/2607.10059"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2607-10059", "source": "claire_radar", "source_benchmark_id": "2607.10059", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "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": 8, "hfPaperUrl": "https://huggingface.co/papers/2607.10059", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"state": "not_applicable"}, "hfDatasetDownloads": {"state": "not_applicable"}, "hfDatasetLikes": {"state": "not_applicable"}, "hfPaperUpvotes": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}}, "status": "fresh"}, "capabilities": ["Reasoning"], "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": "72e3cdd0dcd558b7f52a77710bf5c5d8655a5f4569db27a6fc48214e0bed26dd", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "AgentAbstain is a paired-task benchmark for evaluating LLM agents' ability to abstain from acting in scenarios such as ambiguity, conflicting constraints, or tool failures. It includes 263 paired tasks across 42 sandbox environments, with a proposed pipeline for generating fresh task instances.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "evaluation protocol evidence"], "snippet": "At its core, AgentAbstain is a paired-task benchmark built on an agent-native taxonomy of 8 abstention scenarios across pre-execution reasoning and runtime discovery."}, "familyId": "bmf_4b74dce329f5", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 0, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [], "status": "not-found"}, "heat": null, "id": "bm_agentabstain_d94f5610", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": null, "data": null, "hfPaper": "https://huggingface.co/papers/2607.10059", "paper": "https://arxiv.org/abs/2607.10059", "pdf": "https://arxiv.org/pdf/2607.10059", "project": null, "report": "https://arxiv.org/abs/2607.10059"}, "metrics": [], "motivation": "Agent systems based on large language models (LLMs) are increasingly deployed for autonomous tasks, yet existing evaluations mostly focus on task success rather than whether agents know when to abstain.", "name": "AgentAbstain", "oneLine": "AgentAbstain is a paired-task benchmark for evaluating LLM agents' ability to abstain from acting in scenarios such as ambiguity, conflicting constraints, or tool failures. It includes 263 paired tasks across 42 sandbox environments, with a proposed pipeline for generating fresh task instances.", "paperTitle": "AgentAbstain: Do LLM Agents Know When Not to Act?", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2607.10059", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "ranking": {"90d": {"confidence": "Low", "coverage": 0.15, "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": 0.3882377386428991, "value": 8}}, "confidence": "Low", "coverage": 0.15, "rank": null, "score": 48}, "method": "attention-ranking-v13", "rank": null, "score": 48}}, "readiness": "Paper only", "recognitionConfidence": 0.85, "relation": "introduces", "releasedAt": "2026-07-11", "source": {"authors": ["Xun Liu", "Yi Evie Zhang", "Vira Kasprova", "Parisa Rabbani", "Pardis Sadat Zahraei", "Tianyu Zhang", "Ali Ebrahimpour-Boroojeny", "Varun Chandrasekaran"], "categories": ["cs.AI"], "id": "2607.10059", "title": "AgentAbstain: Do LLM Agents Know When Not to Act?", "type": "arxiv", "url": "https://arxiv.org/abs/2607.10059"}, "sourceUpdatedAt": "2026-07-11T00:57:36Z", "topics": ["Agents", "Reasoning"], "venueAttempts": [], "whyItMatters": "Agent abstention is critical for safe deployment, yet existing evaluations focus on task success. This benchmark targets the gap in measuring calibrated abstention, highlighting that abstention capability is independent of general task-solving ability."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
