{"record": {"aliases": [], "artifacts": [{"id": "arxiv:2606.22723", "kind": "paper", "url": "https://arxiv.org/abs/2606.22723"}, {"id": "gh:tropicai-research/bluexv2", "kind": "repo", "url": "https://github.com/TropicAI-Research/BLUEXv2"}, {"id": "hf:tropic-ai/bluex-v2", "kind": "dataset", "url": "https://huggingface.co/datasets/Tropic-AI/BLUEX-v2"}], "authors": ["João Guilherme Alves Santos", "Giovana Kerche Bonás", "Thiago Laitz", "Thales Sales Almeida", "Helio Pedrini"], "authors_reference": {"basis": "arxiv_metadata", "retrieved_at": "2026-11-03", "source_key": "claire-radar:2606.22723", "source_url": "https://arxiv.org/abs/2606.22723"}, "categories": ["Language & Knowledge", "Mathematics & Formal Science", "cs.CL"], "description": {"en": "BLUEX v2 evaluates large language models on open-ended, discursive questions from the second-phase entrance exams of UNICAMP and USP (2022–2025). The dataset includes 395 questions with 919 subquestions, covering nine subjects, with 55.7% of questions containing images represented as context-aware captions. Scoring uses an LLM-as-a-judge protocol with binary rubric criteria based on official reference answers, yielding a 0–10 score."}, "documents": [{"document_type": "registry_page", "id": "claire_radar:https://arxiv.org/abs/2606.22723", "source": "claire_radar", "source_url": "https://arxiv.org/abs/2606.22723", "title": "BLUEX v2"}], "evidence_summary": {"document_count": 1, "model_count": null, "model_count_basis": null}, "key": "claire-radar:2606.22723", "modality": null, "name": "BLUEX v2", "openness": {"code_license": null, "data_license": null, "evidence": [], "status": "unknown"}, "provenance": {"confidence": "High", "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_bluex-v2_e4bf2712", "origin_source": "arxiv", "recognition_confidence": "1.0", "record_sha256": "c0e0cdb0d9f7bc0a6cbfcd27b2c3c51d50f9a57b9c7183e547ae489fff927a15", "relation": "introduces", "review_state": "unreviewed", "source_url": "https://arxiv.org/abs/2606.22723"}, "publisher": null, "released": "2026-06-21", "released_reference": {"basis": "benchmark_release", "source_key": "claire-radar:2606.22723", "source_url": "https://arxiv.org/abs/2606.22723"}, "schema_version": 1, "sizes": [], "slug": "claire-radar-2606-22723", "source": "claire_radar", "source_benchmark_id": "2606.22723", "source_metadata": {"claire_radar": {"adoption30d": null, "aliases": [], "annotation": "Unknown", "applicationDomains": ["Mathematics & Formal Science"], "area": "Language & Knowledge", "attention": {"asOf": "2026-10-05", "attemptedAt": "2026-10-04T15:31:48.030916Z", "githubRepo": "https://github.com/TropicAI-Research/BLUEXv2", "githubScope": "benchmark_repo", "githubStars": 1, "hfDailySubmittedAt": null, "hfDataset": "Tropic-AI/BLUEX-v2", "hfDatasetDownloads": 65, "hfDatasetLikes": 0, "hfPaperUpvotes": 0, "hfPaperUrl": "https://huggingface.co/papers/2606.22723", "observedAt": "2026-10-04T15:31:48.030916Z", "signalStatus": {"githubStars": {"observedAt": "2026-10-04T15:31:48.030916Z", "state": "fresh"}, "hfDatasetDownloads": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}, "hfDatasetLikes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}, "hfPaperUpvotes": {"checkedAt": "2026-10-04T15:31:48.030916Z", "lastSuccessfulDate": "2026-10-04", "state": "stale"}}, "status": "fresh"}, "capabilities": [], "confidence": "High", "construction": "Transform Existing", "constructionDetail": "Unknown — the indexer does not infer construction details without explicit source evidence.", "copyGeneration": {"generatedAt": "2026-08-20T15:22:10.044731Z", "inputHash": "25b9bea1583f3f0749a1aec78df6c8536a7f11a287be1b637f2ee899196b0067", "model": "deepseek-v4-flash"}, "dataStatus": "primary-source-indexed", "demo": false, "description": "BLUEX v2 evaluates large language models on open-ended, discursive questions from the second-phase entrance exams of UNICAMP and USP (2022–2025). The dataset includes 395 questions with 919 subquestions, covering nine subjects, with 55.7% of questions containing images represented as context-aware captions. Scoring uses an LLM-as-a-judge protocol with binary rubric criteria based on official reference answers, yielding a 0–10 score.", "domainCuration": {"method": "rules-v1", "state": "auto"}, "evidence": {"reasonCodes": ["exact coined title identity tied to benchmark evidence", "benchmark release stated in one sentence", "evaluation protocol evidence", "public artifact URL"], "snippet": "In this work, we introduce BLUEX v2, a benchmark derived from the second-phase entrance exams of Brazil's two leading universities: UNICAMP (Comvest) and USP (Fuvest), spanning exam years 2022--2025."}, "familyId": "bmf_c1d585d91de1", "firstSeenAt": "2026-08-19", "githubIndex": {"candidateCount": 1, "indexedAt": "2026-09-17T04:40:34.111665+00:00", "repositories": [{"evidence": [{"basis": "existing-code-link", "linkedUrl": "https://github.com/TropicAI-Research/BLUEXv2", "url": "https://github.com/TropicAI-Research/BLUEXv2"}, {"basis": "existing-metric-association", "linkedUrl": "https://github.com/TropicAI-Research/BLUEXv2", "url": "https://huggingface.co/papers/2606.22723"}, {"basis": "linked-from-report", "linkedUrl": "https://github.com/TropicAI-Research/BLUEXv2", "url": "https://arxiv.org/abs/2606.22723"}], "scope": "benchmark_repo", "url": "https://github.com/TropicAI-Research/BLUEXv2", "verification": "source-linked"}], "status": "linked"}, "heat": null, "id": "bm_bluex-v2_e4bf2712", "indexedAt": "2026-08-19T11:05:13.395391Z", "industrySectors": [], "links": {"code": "https://github.com/TropicAI-Research/BLUEXv2", "data": "https://huggingface.co/datasets/Tropic-AI/BLUEX-v2", "hfPaper": "https://huggingface.co/papers/2606.22723", "paper": "https://arxiv.org/abs/2606.22723", "pdf": "https://arxiv.org/pdf/2606.22723", "project": null, "report": "https://arxiv.org/abs/2606.22723"}, "metrics": [], "motivation": "Although Large Language Models (LLMs) excel in many tasks, their assessment in Portuguese has received less attention, particularly for open-ended, discursive tasks that demand deeper reasoning and generation capabilities.", "name": "BLUEX v2", "oneLine": "BLUEX v2 evaluates large language models on open-ended, discursive questions from the second-phase entrance exams of UNICAMP and USP (2022–2025). The dataset includes 395 questions with 919 subquestions, covering nine subjects, with 55.7% of questions containing images represented as context-aware captions. Scoring uses an LLM-as-a-judge protocol with binary rubric criteria based on official reference answers, yielding a 0–10 score.", "paperTitle": "BLUEX v2: Benchmarking LLMs on Open-Ended Questions from Brazilian University Entrance Exams", "primaryDomain": "Mathematics & Formal Science", "publication": {"evidence": null, "evidenceLevel": "unverified", "evidenceUrl": "https://arxiv.org/abs/2606.22723", "source": "arxiv-metadata", "status": "unverified", "venue": null, "verifiedAt": "2026-10-14T01:46:58.058199Z"}, "publications": [], "publishers": [{"name": "TropicAI Research", "organizationType": "academic-lab", "role": "benchmark-publisher", "sourceUrl": "https://github.com/TropicAI-Research/BLUEXv2"}], "ranking": {}, "readiness": "Runnable", "recognitionConfidence": 1.0, "relation": "introduces", "releasedAt": "2026-06-21", "source": {"authors": ["Jo\\~ao Guilherme Alves Santos", "Giovana Kerche Bon\\'as", "Thiago Laitz", "Thales Sales Almeida", "Helio Pedrini"], "categories": ["cs.CL"], "id": "2606.22723", "title": "BLUEX v2: Benchmarking LLMs on Open-Ended Questions from Brazilian University Entrance Exams", "type": "arxiv", "url": "https://arxiv.org/abs/2606.22723"}, "sourceUpdatedAt": "2026-06-30T12:58:36Z", "topics": ["cs.CL"], "venueAttempts": [], "whyItMatters": "Portuguese-language evaluation of LLMs has been limited, especially for open-ended tasks requiring deep reasoning and generation. This benchmark provides a public, reusable testbed for assessing capabilities in mathematical reasoning, image understanding, and other dimensions, offering comparable scores across models for practical model selection."}}}, "schema_version": 1, "scores_by_source": {}, "siblings": []}
