AI can read an image yet struggle to reason about it

The same question, a different result

Multimodal language models process both words and images, but their problem-solving ability can depend on how information arrives. Gregory Kang Ruey Lau, Minh Huynh Nguyen and Bryan Kian Hsiang Low investigate this discrepancy in an arXiv preprint posted on November 7, 2026. This briefing draws on the supplied paper text, rather than the abstract alone.

The researchers turned text-based reasoning questions into images, initially using black lettering on a white background. That preserved the question’s content while changing its input format. They then compared three conditions: solving the original text, solving its image, and solving text that the model itself had transcribed from that image.

The third condition matters. If a model cannot read a question, poor answers are unsurprising. But if its transcription supports a correct answer while the image does not, recognition errors cannot fully explain the difference.

Across four initial models, images consistently produced lower accuracy than original text on the MATH-500 mathematics benchmark. For Qwen2.5-VL-3B, accuracy was 50.87% with text, 39.93% with images and 52.93% with extracted text. These results were averaged over three runs.

The pattern was not confined to one rendering style: square layouts and inverted colors also showed gaps. The authors call the discrepancy a reasoning-access gap—visual input appears less effective at engaging capabilities demonstrated with text.

Training on the mismatch

Their proposed approach, TIGER, begins with text mathematics problems whose answers can be automatically checked. Each model receives both text and image versions. The researchers retain cases where it answers correctly from text but incorrectly from the image, creating a training set tailored to that model’s failures.

For the four main models, these sets contained between 1,586 and 4,043 examples. The images were synthetic renderings of text, not collections of manually annotated diagrams or photographs.

TIGER’s main training procedure uses Group Relative Policy Optimization, or GRPO. For each image problem, the model generates several candidate responses. Training favors responses earning higher rewards, primarily for a correct final answer, with a smaller reward for following the required answer format. It does not explicitly reward transcription quality.

The researchers also tested supervised fine-tuning, which imitates successful text-generated solutions, and preference training, which favors those solutions over unsuccessful image-generated responses. A further control applied GRPO to the same problems as text rather than images.

Improvements beyond images of words

Image-conditioned GRPO produced the strongest overall results among these variants. For Qwen3-VL-8B, accuracy on rendered MATH-500 questions increased from 70.9% to 82.4%. That exceeded the original model’s text score, although it should not be confused with a comparison against the newly trained model’s text performance.

More importantly, improvements transferred to benchmarks containing visual material unlike the training images. On MathVerse, which includes mathematical diagrams, Qwen3-VL-8B rose from 54.5% to 65.8%—an increase of 11.3 percentage points. Its EMMA score, covering multidisciplinary reasoning, increased from 30.6% to 36.3%.

Gains on broader evaluations were generally smaller. Across MMStar, ChartQAPro, MMMU validation and RoboSpatial-Home, the four main models’ average improvements ranged from 0.9 to 2.3 percentage points. These tests cover general visual understanding, charts, multidisciplinary questions and spatial relationships.

Training on images outperformed training on the same problems as text overall. Targeted examples also beat equally sized random training subsets in the reported InternVL comparison. TIGER additionally improved the average benchmark performance of several existing visual-reasoning training approaches, though not every individual score increased.

What the evidence does—and does not—explain

The authors examined internal activation patterns associated with text-side reasoning. After image-conditioned GRPO, visual inputs aligned more strongly with these patterns. Transferring intermediate visual representations from trained models into an unchanged base model also improved accuracy. These interventions support the proposed explanation, but do not identify a complete or unique reasoning mechanism.

There are important boundaries. Successful transcription in a separate task does not establish that the model represents every relevant detail identically while solving an image directly. The experiments show that perception failures alone are insufficient, not that perception no longer matters.

Training was limited to mathematics-derived text images, which underrepresent ambiguous real-world visuals. Benchmark improvements do not establish deployment reliability. Training schedules also differed between methods; extending the comparison methods did not improve their average results in the reported checks, but this was not a fully compute-matched comparison.

The central result is therefore specific: training on text-success/image-failure cases improved visual benchmark performance. The broader interpretation—that visual inputs can fail to engage otherwise available reasoning capabilities—is supported, rather than conclusively proved.