{"record": {"aliases": ["Verus-SpecGym: An Agentic Environment for Evaluating Specification Autoformalization"], "artifacts": [{"id": "arxiv:2605.26457", "kind": "paper", "url": "https://arxiv.org/abs/2605.26457"}, {"id": "gh:formal-verif-is-cool/verus-spec-gym", "kind": "repo", "url": "https://github.com/formal-verif-is-cool/verus-spec-gym"}], "authors": ["Anmol Agarwal", "Natalie Neamtu", "Pranjal Aggarwal", "Seungone Kim", "Jannis Limperg", "Cedric Flamant", "Kanna Shimizu", "Bryan Parno", "Sean Welleck"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2605.26457", "source_url": "https://arxiv.org/abs/2605.26457"}, "categories": ["Code & Software", "General AI", "cs.SE"], "description": {"en": "Verus-SpecBench has 581 tasks in which agents write formal Verus specifications for Codeforces problems. Specs are executed and tested against official tests and hack cases."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2605.26457", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2605.26457", "title": "Verus-SpecBench"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "institutions": ["Carnegie Mellon University", "Amazon (United States)"], "institutions_reference": {"basis": "openalex_linked_institutions", "doi": "10.48550/arxiv.2605.26457", "retrieved_at": "2026-11-04", "source_key": "claire-radar:2605.26457", "source_url": "https://openalex.org/W7162515531"}, "key": "claire-radar:2605.26457", "modality": null, "name": "Verus-SpecBench", "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_verus-specgym-an-agentic-environment-for-e_d9a9ebc1", "origin_source": "arxiv", "recognition_confidence": "0.7", "record_sha256": "de834949311492e93a4f1a5c298ee4da08dec00728a5ab90707b69c632ccad4c", "relation": "introduces", "review_model": "claude-code-default", "review_state": "ai-reviewed", "reviewed_at": "2026-09-24T16:47:06.238906Z", "source_url": "https://arxiv.org/abs/2605.26457"}, "publisher": null, "released": "2026-05-26", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2605.26457", "source_url": "https://arxiv.org/abs/2605.26457"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2605-26457", "source": "claire_radar", "source_benchmark_id": "2605.26457", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": ["Verus-SpecGym: An Agentic Environment for Evaluating Specification Autoformalization"], "annotation": "Unknown", "applicationDomains": ["General AI"], "area": "Code & Software", "attentionForecast": {"confidence": "Medium", "horizon": "7d", "reason": "Formal verification for coding agents is an active area. Frontier model results and a public environment are likely to draw attention.", "score": 55}, "capabilities": [], "confidence": "Medium", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-09-24T16:47:06.238906Z", "inputHash": "9c01ba91f9dbfb0ac7fa69605252ae09a5822f8e06500f9b3eed0e28d9d27c98", "model": "claude-code-default"}, "curation": {"admissionPolicyVersion": "2026-09-09.1", "canonicalNameEvidence": "We introduce Verus-SpecBench, a benchmark of 581 spec-writing tasks", "canonicalNameSource": "abstract", "decisionReason": "It is formally named and uses a fixed task set. Scoring is repeatable, based on executing specs against tests, and reports task solve rates. Code, data, and logs are public.", "evaluationEvidence": {"reuse": "The code, data, & logs can be found at https://github.com/formal-verif-is-cool/verus-spec-gym", "scoring": "testing them against official Codeforces tests & adversarial cases extracted from Codeforces \"hacks\"", "task": "a benchmark of 581 spec-writing tasks derived from Codeforces problems targeting Verus, a verifier for Rust"}, "model": "claude-code-default", "reviewedAt": "2026-09-24T16:47:06.238906Z", "state": "ai-reviewed"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "Verus-SpecBench has 581 tasks in which agents write formal Verus specifications for Codeforces problems. Specs are executed and tested against official tests and hack cases.", "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 Verus-SpecBench, a benchmark of 581 spec-writing tasks derived from Codeforces problems targeting Verus, a verifier for Rust, and Verus-SpecGym, an agentic environment in which models interact with Verus, bash, & the filesystem to develop these specs."}, "familyId": "bmf_3474a5ece488", "firstSeenAt": "2026-08-19", "heat": null, "id": "bm_verus-specgym-an-agentic-environment-for-e_d9a9ebc1", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/formal-verif-is-cool/verus-spec-gym", "data": null, "paper": "https://arxiv.org/abs/2605.26457", "pdf": "https://arxiv.org/pdf/2605.26457", "project": null, "report": "https://arxiv.org/abs/2605.26457"}, "metrics": [], "motivation": "AI coding agents are increasingly used to write real-world software, but ensuring that their outputs are correct remains a fundamental challenge.", "name": "Verus-SpecBench", "oneLine": "Verus-SpecBench has 581 tasks in which agents write formal Verus specifications for Codeforces problems. Specs are executed and tested against official tests and hack cases.", "paperTitle": "Verus-SpecGym: An Agentic Environment for Evaluating Specification Autoformalization", "primaryDomain": "General AI", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2605.26457", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-08-19T11:05:13.395391Z"}, "publications": [], "readiness": "Runnable", "recognitionConfidence": 0.7, "relation": "introduces", "releasedAt": "2026-05-26", "reviewLifecycle": {"state": "pending"}, "source": {"authors": ["Anmol Agarwal", "Natalie Neamtu", "Pranjal Aggarwal", "Seungone Kim", "Jannis Limperg", "Cedric Flamant", "Kanna Shimizu", "Bryan Parno", "Sean Welleck"], "categories": ["cs.SE", "cs.AI", "cs.CL", "cs.PL"], "id": "2605.26457", "title": "Verus-SpecGym: An Agentic Environment for Evaluating Specification Autoformalization", "type": "arxiv", "url": "https://arxiv.org/abs/2605.26457"}, "sourceUpdatedAt": "2026-05-26T02:12:48Z", "topics": ["cs.SE"], "venueAttempts": [], "whyItMatters": "It tests whether agents can turn informal requirements into faithful formal specifications. The scoring is execution-based, and the paper reports that LLM judges miss 26% of the failures it catches."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
