{"paper":{"title":"AGIEval: A Human-Centric Benchmark for Evaluating Foundation Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"AGIEval benchmark shows GPT-4 surpassing average humans on SAT math at 95 percent and LSAT.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Amin Saied, Nan Duan, Ruixiang Cui, Shuai Lu, Wanjun Zhong, Weizhu Chen, Yanlin Wang, Yaobo Liang, Yiduo Guo","submitted_at":"2023-04-13T09:39:30Z","abstract_excerpt":"Evaluating the general abilities of foundation models to tackle human-level tasks is a vital aspect of their development and application in the pursuit of Artificial General Intelligence (AGI). Traditional benchmarks, which rely on artificial datasets, may not accurately represent human-level capabilities. In this paper, we introduce AGIEval, a novel benchmark specifically designed to assess foundation model in the context of human-centric standardized exams, such as college entrance exams, law school admission tests, math competitions, and lawyer qualification tests. We evaluate several state"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"GPT-4 surpasses average human performance on SAT, LSAT, and math competitions, attaining a 95% accuracy rate on the SAT Math test and a 92.5% accuracy on the English test of the Chinese national college entrance exam.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That standardized human exams are valid and representative proxies for general human-level cognitive capabilities without significant selection bias or format advantages for current model architectures.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"AGIEval shows GPT-4 exceeding average human scores on SAT Math at 95% and Chinese college entrance English at 92.5%, while revealing weaker results on complex reasoning tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"AGIEval benchmark shows GPT-4 surpassing average humans on SAT math at 95 percent and LSAT.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d7b2d23d251b913928ce3f45aa96b82d7a80957aa443931c8947354c3da3b233"},"source":{"id":"2304.06364","kind":"arxiv","version":2},"verdict":{"id":"35bdc164-9f80-46b8-8082-15b60c51cf65","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T09:59:54.385025Z","strongest_claim":"GPT-4 surpasses average human performance on SAT, LSAT, and math competitions, attaining a 95% accuracy rate on the SAT Math test and a 92.5% accuracy on the English test of the Chinese national college entrance exam.","one_line_summary":"AGIEval shows GPT-4 exceeding average human scores on SAT Math at 95% and Chinese college entrance English at 92.5%, while revealing weaker results on complex reasoning tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That standardized human exams are valid and representative proxies for general human-level cognitive capabilities without significant selection bias or format advantages for current model architectures.","pith_extraction_headline":"AGIEval benchmark shows GPT-4 surpassing average humans on SAT math at 95 percent and LSAT."},"references":{"count":287,"sample":[{"doi":"10.24963/ijcai.2022/629","year":2022,"title":"Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence,","work_id":"c4f3637d-3692-46a9-bfea-c7bd6b57ae7f","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages=","work_id":"c6b78093-c72e-47cb-afc3-1e6df42e13ca","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"2023 , publisher =","work_id":"ec5f4dd4-c17c-4ada-92fd-e4e521e15710","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Communications of the ACM , volume=","work_id":"d5c1eec5-fe7c-47e3-b110-9b2ca7817689","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=","work_id":"83cfc672-4920-4faa-8e23-5c30b4777e6f","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":287,"snapshot_sha256":"82933c7ffcb95de91c268d61a400ba36d7f656730ad69d223ad26b41a33716bc","internal_anchors":12},"formal_canon":{"evidence_count":2,"snapshot_sha256":"0957bd48e13f7caae39084c23b61c2e1a362a81a1c84c48e796139f1b6f376fa"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}