Seeing isn't reasoning: TIGER trains only on rendered text images, lifting Qwen3-VL-8B on MathVerse from 54.5 to 65.8
Synopsis
Multimodal large language models can read text in an image accurately yet still fail to reason over the same problem as well as they do over plain text, a shortfall the authors call a modality reasoning-access gap. TIGER renders text-only math problems as images, mines cases the model solves from text but fails from the image, and post-trains on those image inputs with GRPO and related objectives. Trained only on synthetic rendered-text images, Qwen3-VL-8B rises from 54.5 to 65.8 on MathVerse, while Qwen2.5-VL-7B rises from 43.9 to 47.5 on MathVerse and from 22.2 to 25.7 on EMMA.
Interpretation
On semantically identical problems, multimodal large language models score higher from text than from a rendered image of the same problem, and this drop cannot be attributed to recognition failure alone. The prevailing explanation treated multimodal reasoning failure mainly as a perception problem, assuming that correctly extracted visual content would be reasoned over as effectively as text. On MATH-500 and its black-text-on-white rendering MATH-500-I, image input lowers accuracy by 7.2 points on average, with the largest drop for Qwen2.5-VL-3B; among incorrect image-based answers, 92.36% occur even when the problem text has been extracted correctly.
TIGER turns the model's own text-success, image-failure cases into training signal so that image input engages reasoning the model already has. Training needs no human-annotated image-text reasoning pairs; it starts from text-only math problems with verifiable answers. About 45K solvable problems are retained from OpenR1-Math-220 and then filtered per model, giving training sets of 1586 to 4043 examples across backbones; evaluation covers MATH-500-I, BBH-I and the held-out real benchmarks MathVista, MathVerse, MathVision and EMMA.
Image-conditioned GRPO on the mined examples gives the most consistent gains among the variants and stacks with existing RLVR visual reasoning methods. SFT and DPO variants on the same examples often improve synthetic settings but do not transfer reliably to real multimodal benchmarks; for Qwen3-VL-8B, SFT and DPO improve MathVerse but degrade MathVision and EMMA. Qwen2.5-VL-7B rises from 43.9 to 47.5 on MathVerse and from 22.2 to 25.7 on EMMA, and Qwen3-VL-8B rises from 54.5 to 65.8 on MathVerse; across MMStar, MMMU, ChartQAPro and RoboSpatial-Home the average gain is 0.9 to 2.3 points, and a one-sided Wilcoxon signed-rank test over model-dataset pairs on eight real benchmarks gives p = 4e-7.
Perspective
The demonstrated scope is training data drawn entirely from text-only math problems in OpenR1-Math-220 and their synthetic rendered images, with evaluation on synthetic benchmarks such as MATH-500-I and BBH-I and on real benchmarks including MathVista, MathVerse, MathVision, EMMA, MMStar, MMMU, ChartQAPro and RoboSpatial-Home, across backbones including InternVL3.5-1B, Qwen2.5-VL-3B, Qwen2.5-VL-7B, Qwen3-VL-8B and LLaVA-OneVision-1.5-4B. The authors propose extending the framework to real-world multimodal tasks and investigating mechanisms that more tightly couple visual encoding with reasoning.
The authors note that the method is constrained by the domain coverage of the source text corpora and that synthetic image construction may underrepresent the ambiguity of in-the-wild visuals; the reasoning-subspace analysis yields an empirical subspace associated with text-side reasoning rather than a causal identification of the reasoning mechanism.