{"record": {"aliases": ["When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents"], "artifacts": [{"id": "arxiv:2606.05806", "kind": "paper", "url": "https://arxiv.org/abs/2606.05806"}, {"id": "gh:zhudongsheng75/toolmaze", "kind": "repo", "url": "https://github.com/Zhudongsheng75/ToolMaze"}], "authors": ["Dongsheng Zhu", "Xuchen Ma", "Yucheng Shen", "Xiang Li", "Yukun Zhao", "Shuaiqiang Wang", "Lingyong Yan", "Dawei Yin"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.05806", "source_url": "https://arxiv.org/abs/2606.05806"}, "categories": ["General AI", "Language & Knowledge", "cs.AI"], "description": {"en": "A benchmark that tests how tool-using LLM agents replan and recover from tool failures. It varies task structure (DAG-based complexity) and failure type (explicit/implicit, transient/permanent)."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.05806", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.05806", "title": "ToolMaze"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Shanghai Artificial Intelligence Laboratory", "East China Normal University", "Soochow University", "Shandong University", "Baidu (China)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2606.05806", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2606.05806", "source_url": "https://openalex.org/W7163681859"}, "key": "claire-radar:2606.05806", "modality": null, "name": "ToolMaze", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"admission_policy_version": "2026-09-09.1", "confidence": "Medium", "crawl_bundle": "claire_radar_2026-10-14", "crawled_at": "2026-09-25T04:40:22Z", "data_status": "primary-source-indexed", "display_eligible": "true", "origin_record_id": "bm_when-tools-fail-benchmarking-dynamic-repla_600aa235", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "396a0d5183f69302f17acae2867f6389cd1f12a8060dff13c7031d00c59719bb", "relation": "introduces", "review_model": "claude-code-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T16:55:21.506802Z", "source_url": "https://arxiv.org/abs/2606.05806"}, "publisher": null, "released": "2026-06-04", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.05806", "source_url": "https://arxiv.org/abs/2606.05806"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-05806", "source": "claire_radar", "source_benchmark_id": "2606.05806", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Language & Knowledge", "attentionForecast": {"confidence": "Low", "horizon": "7d", "reason": "Tool-failure robustness is a topic of active interest for agent research, but the benchmark has no project site or institutional backing in the source text.", "score": 40}, "capabilities": [], "confidence": "Medium", "construction": "Unknown", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T16:55:21.506802Z", "inputHash": "0f624a22560a73813fcb9e6d5792d93f511d839536e23feb8be37c55fbb4bab7", "model": "claude-code-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce ToolMaze, a benchmark", "canonicalNameSource": "abstract", "decisionReason": "It has a formal name, a defined task design, a named recovery metric (PRR), and public data and code.", "evaluationEvidence": {"reuse": "Data and code are available at https://github.com/Zhudongsheng75/ToolMaze", "scoring": "Perturbation Recovery Rate (PRR)", "task": "a benchmark for dynamic path discovery and error recovery in TIR agents"}, "model": "claude-code-default", "reviewedAt": "2026-09-24T16:55:21.506802Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "A benchmark that tests how tool-using LLM agents replan and recover from tool failures. It varies task structure (DAG-based complexity) and failure type (explicit/implicit, transient/permanent).", "displayEligible": true, "domainCuration": {"method": "rules-v1", "state": "auto"}, "evaluationMode": "public_reusable", "evidence": {"reasonCodes": ["benchmark term in abstract", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "We introduce ToolMaze, a benchmark for dynamic path discovery and error recovery in TIR agents."}, "familyId": "bmf_be40d4653dd1", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_when-tools-fail-benchmarking-dynamic-repla_600aa235", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/Zhudongsheng75/ToolMaze", "data": null, "paper": "https://arxiv.org/abs/2606.05806", "pdf": "https://arxiv.org/pdf/2606.05806", "project": null, "report": "https://arxiv.org/abs/2606.05806"}, "metrics": [], "motivation": "Existing benchmarks evaluate Tool-Integrated Reasoning (TIR) in LLMs on idealized ''happy paths'', largely overlooking real-world tool failures.", "name": "ToolMaze", "oneLine": "A benchmark that tests how tool-using LLM agents replan and recover from tool failures. It varies task structure (DAG-based complexity) and failure type (explicit/implicit, transient/permanent).", "paperTitle": "When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.05806", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-06-04", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Dongsheng Zhu", "Xuchen Ma", "Yucheng Shen", "Xiang Li", "Yukun Zhao", "Shuaiqiang Wang", "Lingyong Yan", "Dawei Yin"], "categories": ["cs.AI"], "id": "2606.05806", "title": "When Tools Fail: Benchmarking Dynamic Replanning and Anomaly Recovery in LLM Agents", "type": "arxiv", "url": "https://arxiv.org/abs/2606.05806"}, "sourceUpdatedAt": "2026-06-04T07:38:46Z", "topics": ["cs.AI"], "venueAttempts": [], "whyItMatters": "It goes beyond scenarios where tools always work, and reports that fault tolerance improves with model scale 3.66x more slowly than basic task execution."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
