To prevent agents from exploiting shallow heuristics from titles or abstracts, SciResearchBench adopts afull-text-firstconstruction methodology:
•Broad Academic Scope: 1,000 expert-curated, challenging instances (600 Deep, 400 Wide) spanning 8 core computer science areas, including CV, LLMs, Theory, and AI4Science.
•Grounded, Obfuscated Clues: Evidence is buried in methodology details, empirical tables, figure captions, ablation studies, and appendices, with surface terms obfuscated to eliminate lexical shortcuts.
•Human-in-the-Loop Quality Control: Verified across multiple independent rounds by CS PhD annotators, multi-model consensus voting, and adversarial stress tests, achieving a 75.9% candidate rejection rate to ensure gold-standard rigor.