Auditable Step Verification for Vision-Language Reasoning

Outcome-only reinforcement learning from verifiable rewards cannot tell whether a vision-language model failed because it misread the image or because it reasoned incorrectly after reading it. We present $\textbf{VisualSPARK}$, a generative process reward model for multimodal math reasoning. VisualSPARK learns from Monte-Carlo-labeled reasoning steps and produces an auditable verification rationale ending in a $\texttt{[[correct]]}$ or $\texttt{[[incorrect]]}$ verdict. Across MathVista, MathVerse, MathVision, and MMMU-Pro, we evaluate on $77{,}680$ cached completions. The verifier reaches $81.3\%$ macro-F1 with $100\%$ parseable held-out verdicts. On $7{,}980$ free-form problems, PRM-weighted majority adds $+5.26$ aggregate points over pass@$1$---about $420$ expected additional correct selections---while preserving $98.6\%$ of majority-vote lift; pass@$8$ reveals $+25.57$ points of recoverable headroom. Diagnostics rule out length as a within-problem shortcut, quantify contamination and extraction sensitivity, and show practical batched serving. We release the full code, configs, result logs, datasheets, model cards, and measured compute accounting.