SeePhys Pro: Diagnosing Modality Transfer and Blind-Training Effects in Multimodal RLVR for Physics Reasoning

TL;DR

SeePhys Pro introduces a benchmark for evaluating how well multimodal models maintain reasoning capabilities when information shifts from text to images, revealing current models' limitations and the effects of blind training.

Contribution

The paper presents a new modality transfer benchmark and analyzes the impact of blind training on multimodal reasoning robustness.

Findings

Models' performance drops as information moves from language to diagrams.

Blind training with masked images can improve performance without visual evidence.

Residual cues, not visual evidence, may drive some performance gains.

Abstract

We introduce SeePhys Pro, a fine-grained modality transfer benchmark that studies whether models preserve the same reasoning capability when critical information is progressively transferred from text to image. Unlike standard vision-essential benchmarks that evaluate a single input form, SeePhys Pro features four semantically aligned variants for each problem with progressively increasing visual elements. Our evaluation shows that current frontier models are far from representation-invariant reasoners: performance degrades on average as information moves from language to diagrams, with visual variable grounding as the most critical bottleneck. Motivated by this inference-time fragility, we further develop large training corpora for multimodal RLVR and use blind training as a diagnostic control, finding that RL with all training images masked can still improve performance on unmasked validation sets. To analyze this effect, text-deletion, image-mask-rate, and format-saturation controls suggest that such gains can arise from residual textual and distributional cues rather than valid visual evidence. Our results highlight the need to evaluate multimodal reasoning not only by final-answer accuracy, but also by robustness under modality transfer and by diagnostics that test whether improvements rely on task-critical visual evidence.