Microsoft researchers introduce SciForma, a framework for generating scientific methodology diagrams with explicit structural fidelity. It decomposes quality into three axes: components, arrows, and text, supported by a structural inventory. The work presents SciFormaData-700K for training and SciFormaBench-2K for logic-verified evaluation, then proposes Multi-Dimensional Conjunctive Preference Optimization (M-DPO), which routes optimization toward the weakest structural dimension while requiring simultaneous correctness across all axes. The authors report that SciForma-9B surpasses open-source baselines and GPT-Image-1.5 on SciFormaBench-2K and AIBench. Code and data are announced for release.
No heat snapshots are available in the last 24 hours.