SciVQR: A Multidisciplinary Multimodal Benchmark for Advanced Scientific Reasoning Evaluation

Longteng Guo, Xuanxu Lin, Dongze Hao, Tongtian Yue, Pengkang Huo, Jiatong Ma, Yuchen Liu, Jing Liu · 2026

Read SciVQR: A Multidisciplinary Multimodal Benchmark for Advanced Scientific Reasoning Evaluation by Longteng Guo, Xuanxu Lin, Dongze Hao, Tongtian Yue, Pengkang Huo, Jiatong Ma, Yuchen Liu, Jing Liu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-10187.txt.

Description

Guo and seven co-authors introduce SciVQR, a cross-field multi-sense benchmark for advanced scientific reasoning. Scientific reasoning blends multi-sense inputs, domain skill and multi-step reasoning, but current benchmarks for multi-sense language models miss the depth and audit trail that strict grading needs. SciVQR spans fifty-four subfields across mathematics, physics, chemistry, geography, astronomy and biology, with domain visuals like equations, charts and diagrams.

Tasks run from basic factual recall to complex multi-step reasoning, with forty-six percent carrying expert-written solutions. Crucially, SciVQR grades the reasoning process as well as final answers, revealing how models reach answers. Evaluations of leading closed and open-source models expose large limits on complex multi-sense reasoning, pointing to the need for better multi-step reasoning and deeper cross-field mixing before models approach true scientific intelligence. Data and grading code are public.

Catalog record

CREATORLongteng Guo, Xuanxu Lin, Dongze Hao, Tongtian Yue, Pengkang Huo, Jiatong Ma, Yuchen Liu, Jing Liu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.10187

Browse the full library · About the Tower of Babel