Science Edge Evaluation: SEE the Missing Step Toward Real Scientific Discovery

Science Edge Evaluation, or SEE, tests whether multimodal large language models can perform evidence-bounded reasoning on authentic laboratory problems rather than merely recall scientific facts. Its benchmark contains 1116 expert-curated questions spanning chemistry, biology, materials science, and 17 fine-grained subfields, using evidence such as NMR and mass spectra, microscopy, Western blots, Cryo-EM, diffraction patterns, and numerical measurements. Across 19 models, including OpenAI’s GPT-5.6-Sol, Google DeepMind’s Gemini 3.1 Pro, Anthropic’s Claude Opus 5, and Alibaba’s Qwen3.8-Max, the best standard accuracy is only 48.7%, while general-purpose models average 34.9% versus 22.2% for science-specialized models. Removing images reveals weak awareness of evidential limits: models explicitly acknowledge missing visual information in just 4.6% of text-only cases. Adding web search and a code interpreter improves the top result to 52.7%, but tool trajectories also create errors through incorrect action selection, visual misinterpretation, and overreliance on retrieved information. SEE’s central finding is that current MLLMs can explain established science yet remain unreliable at extracting precise observations, coordinating evidence across disciplines, maintaining global logical coherence, and limiting conclusions to what experiments actually support. The missing capability for real scientific discovery is therefore not simply more knowledge or tools, but disciplined management of multimodal evidence and uncertainty.