{"schema":"https://assignee.net/schemas/benchmark-evidence-v1","schema_version":"1.0","contract_version":"benchmark-evidence-v1.0","contract_updated":"2026-06-01","schema_documentation":"https://assignee.net/schemas","changelog_url":"https://assignee.net/changelog","publisher":{"name":"Assignee Research","url":"https://assignee.net"},"html_url":"https://assignee.net/benchmarks/evidence?model=GPT-4&bench=MMLU","json_url":"https://assignee.net/benchmarks/evidence.json?model=GPT-4&bench=MMLU","model":"GPT-4","benchmark":"MMLU","source_count":7,"source_coverage":{"record_count":7,"distinct_source_count":7,"coverage_level":"BROAD","basis":"distinct public paper URLs or titles in this evidence cluster"},"source_profile":{"source_url_count":7,"missing_source_url_count":0,"domains":["arxiv.org","doi.org"],"year_min":2023,"year_max":2026,"basis":"public source URLs, source titles, and reported publication years in this evidence cluster"},"reported_range":{"min_score_pct":57.0,"max_score_pct":87.3},"spread_pp":30.3,"severity":"HIGH","entries":[{"model":"GPT-4","benchmark":"MMLU","score_pct":87.3,"source_title":"Adaptive Self-Prompting in Agentic LLM Frameworks for Code Fault Detection","source_url":"https://doi.org/10.3390/software5020016","source_domain":"doi.org","year":2026},{"model":"GPT-4","benchmark":"MMLU","score_pct":87.3,"source_title":"Vendi-RAG: Adaptively Trading-Off Diversity And Quality Significantly Improves Retrieval Augmented Generation With LLMs","source_url":"http://arxiv.org/abs/2502.11228v2","source_domain":"arxiv.org","year":2025},{"model":"GPT-4","benchmark":"MMLU","score_pct":87.3,"source_title":"MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark","source_url":"http://arxiv.org/abs/2406.01574v6","source_domain":"arxiv.org","year":2024},{"model":"GPT-4","benchmark":"MMLU","score_pct":87.3,"source_title":"Understanding the Effectiveness of Large Language Models in Detecting Security Vulnerabilities","source_url":"https://doi.org/10.48550/arxiv.2311.16169","source_domain":"doi.org","year":2023},{"model":"GPT-4","benchmark":"MMLU","score_pct":87.3,"source_title":"Capabilities of GPT-4 on Medical Challenge Problems","source_url":"http://arxiv.org/abs/2303.13375v2","source_domain":"arxiv.org","year":2023},{"model":"GPT-4","benchmark":"MMLU","score_pct":86.4,"source_title":"Data Engineering for Scaling Language Models to 128K Context","source_url":"http://arxiv.org/abs/2402.10171v1","source_domain":"arxiv.org","year":2024},{"model":"GPT-4","benchmark":"MMLU","score_pct":57.0,"source_title":"Investigating Data Contamination in Modern Benchmarks for Large Language Models","source_url":"https://doi.org/10.48550/arxiv.2311.09783","source_domain":"doi.org","year":2023}],"interpretation":"This record groups score claims extracted from papers for the same model and benchmark label. A nonzero spread means the public literature reports different values for this cluster.","limitations":["Differences are not automatically errors.","Reported values may differ because of prompts, dataset versions, evaluation protocols, scoring rules, preprocessing, fine-tuning, or reporting conventions.","Source papers remain authoritative for their own claims."]}