Abstract
Paper abstract
A concise account of the research question, protocol and principal findings.
Open-ended university problems expose whether a model can construct and verify a solution rather than merely recognize an answer. The objective is to compare solution correctness and completeness together with the resources required to obtain them. A controlled, closed-resource benchmark evaluates 14 model-mode configurations on 60 authentic examination problems in Mathematical Analysis III and General Physics II. Sixty item-specific rubrics contain 1,600 atomic criteria; answers are assessed anonymously in single-response contexts, centrally validated, and every case of substantive doubt is escalated to versioned human adjudication.
The six leading configurations occupy a narrow 96.71-98.72% band, and paired bootstrap intervals do not separate the first system from its two nearest competitors. In a same-weights DeepSeek V4 Flash contrast, high reasoning adds 10.93 percentage points (95% CI 7.97-14.10), while increasing token use by 214.7%, cost by 224.4%, and median latency 5.13-fold. Across the normalized matched panel, model-level analytical-token volume, estimated API cost, and median latency vary by factors of 6.91, 234.61, and 12.46. These results show that, once correctness approaches saturation, resource demand, robustness across domains, and quality-resource frontiers become central outcomes.