BAAI and Renmin University publish SciResearchBench, a benchmark for AI literature search

AIBAAI introduces SciResearchBench, a benchmark of 1,000 expert-curated computer science instances spanning eight areas including CV, LLMs, Theory, and AI4Science. Its questions bury evidence in methodology details, tables, figure captions, and appendices, with surface terms obfuscated to block shortcuts from titles or abstracts. Candidates went through multi-round review by CS PhD annotators, multi-model consensus voting, and adversarial tests, with 75.9% rejected.

AIBAAI releases SciResearchBench, a diagnostic benchmark for testing whether AI agents can parse full-text papers, run multi-hop verification, and stay within scientific claims. The dataset, evaluation suite, and DeepXiv interactive environment are open-sourced, with the paper linked on arXiv (2604.25256).

AIBAAI and Renmin University of China say their joint paper introducing SciResearchBench has been accepted to NeurIPS 2026. The authors describe it as the first benchmark for systematically evaluating AI agents on complex, autonomous scientific literature discovery.

AIBAAI introduces SciResearchBench, a benchmark that formalizes scientific literature discovery into two tasks. Deep Research requires finding one unique target paper through multi-step probing and full-text verification, or confirming that no qualifying paper exists. Wide Research requires compiling the complete set of papers meeting strict scientific conditions, averaging 9.23 valid papers per query.