The one-paragraph version
Multimodal Large Language Models can often solve a math or logic problem when it is given as typed text, but fail the exact same problem when it is given as a picture of that text. The authors show this is not just bad eyesight: the models can read the words in the image almost perfectly, and if you feed their own reading back to them as text, they get the right answer again. They call this a reasoning-access gap — the visual information is perceived but does not switch on the reasoning ability the model already has in text. To fix it, they propose TIGER, which automatically finds cases where the model succeeds on text but fails on the matching image, then retrains it to reason directly from the image using trial-and-error learning. Trained only on simple pictures of text math, TIGER improves performance on much harder real visual tasks like geometry diagrams, charts, science questions, and embodied spatial reasoning.
The problem they're solving, and why it matters
The common story for why vision-language models struggle is perception: if only the model could see the diagram or read the chart more clearly, it could then reason as well as a text-only language model. This paper argues that seeing is not enough.
To test this, the authors create what they call modality counterfactuals. That means taking the same problem and presenting it in two forms with identical meaning: once as plain text, and once as an image made by rendering that text in black on white. A concrete example would be the text “Solve 2x + 5 = 11” versus a picture showing those same characters. If perception were the only bottleneck, performance on the two forms should be the same once the model can read the image.
It is not. Across Qwen2.5-VL, Qwen3-VL, InternVL3.5, LLaVA-OneVision, and even larger frontier models like Gemini-3.6-Flash, image versions score several points lower. On MATH-500, a set of competition math problems, the average drop is about 7 points. On the harder graduate-science set GPQA-Diamond, the average drop is about 7.5 points. Yet transcription accuracy, meaning how well the model can copy the text out of the image, stays high at 90 to 97 percent. When the model is allowed to reason over its own transcription instead of directly over the image pixels, accuracy jumps back up to text level.
This matters because most current work tries to fix visual reasoning by improving grounding, object detection, or perception rewards. If the real failure is routing — correctly seen evidence never reaching the language backbone’s existing reasoning machinery — then better eyes alone will not produce better thinkers.
The key idea, in plain words
Think of the model as having two hallways leading to one reasoning room. The text hallway goes straight to the room. The image hallway goes through an extra lobby — a vision encoder and projector — before it reaches the same room.
Training has taught the two hallways to use similar words for similar things, which the authors call semantic alignment. But it has not taught the image hallway how to knock loudly enough to wake up the step-by-step reasoning process inside the room.
TIGER’s key idea is to turn that failure into training signal without hiring humans to label diagrams. Start from cheap text-only reasoning data with known answers. Render each problem as an image. Ask the model to solve both versions. Keep only the revealing cases: it gets the text version right but the image version wrong. Those are proof the capability exists but is not being accessed from vision. Then train specifically on those cases, forcing the model to practice reasoning while looking at the image, until the visual pathway learns to trigger the same reasoning the text pathway already uses.
How it actually works, step by step
First is diagnosis with three conditions. For each problem, the model is tested on the original text input, on the rendered-image input, and on an extracted-text control. That control means: first ask the model to do optical character recognition on the image, then ask it to reason over the text it just extracted. The gap between original text accuracy and extracted-text accuracy is called the perception-control gap, a measure of loss caused by misreading. The remaining drop from extracted-text to direct-image accuracy is the reasoning-access gap, a measure of loss caused by failing to reason from correctly perceived input.
Second is gap-targeted data construction. The source is OpenR1-Math-220, about 45,000 solvable text math problems. Each problem is rendered as an image to make a paired text problem and image problem. The base model answers both with greedy decoding. Each answer is checked as correct or incorrect against the verifiable answer, creating four groups. The authors keep only the group where text is correct and image is incorrect, called D10. They discard cases where both are correct because there is nothing to learn, cases where both are wrong because the model likely lacks the knowledge entirely, and the rare cases where image succeeds but text fails. For each kept example they store the image prompt plus the good text-derived answer and the bad image-derived answer.
Third is gap-targeted training. The main version uses Group Relative Policy Optimization, a form of reinforcement learning with verifiable reward. In plain terms: for one image prompt, the model generates a group of, in this work, eight different attempts. Each attempt gets a reward that is mostly for getting the final boxed answer right and a little for following the required format with boxed output. Within the group, better-than-average attempts are reinforced and worse ones are discouraged. Importantly, no explicit reward is given for reading well, because reading is already good. The optimization pushes the whole image-conditioned reasoning process to be more correct.
For comparison the authors also try two simpler uses of the same pairs. Supervised Fine-Tuning teaches the model to imitate the correct text-derived answer when shown the image. Direct Preference Optimization teaches the model to prefer the correct text-derived answer over its own incorrect image-derived answer. Both reuse fixed answers, while the reinforcement version explores new reasoning paths from images.
What they showed — results, honestly, with limitations
The controlled gap is robust. It appears under different temperatures, different prompts including describing-then-solving, different fonts and color inversions, and square versus wide images. Generation length is similar for text and image inputs, so the gap is not explained by the model simply writing less.
TIGER narrows that gap and transfers beyond its training. On in-distribution synthetic images, gains are large. For example, Qwen2.5-VL-3B goes from 39.9 to 58.8 on image MATH-500, and Qwen3-VL-8B goes from 70.9 to 82.4, now above its own text baseline of 75.5. The reinforcement version is consistently best, beating the imitation and preference versions by about 3 points on average, and beating a text-only reinforcement control that trains on the same problems but as text. That control is important evidence that improvement comes from learning to reason under visual conditioning, not just from seeing more math.
More surprisingly, training only on rendered text math helps on real diagrams and diverse tasks it never saw: MathVista, MathVerse, MathVision, EMMA for multidisciplinary science, plus MMStar, ChartQAPro, MMMU, and RoboSpatial-Home for embodied spatial relations. For instance Qwen3-VL-8B improves from 54.5 to 65.8 on MathVerse and from 30.6 to 36.3 on EMMA. Across eight realistic benchmarks and four model scales, improvements are statistically significant with a reported p-value of 4e-7. TIGER also stacks with existing visual reinforcement methods like Perception-R1, R1-OneVision, and Vision-SR1, adding further gains.
The analysis suggests a mechanism. The authors define Reasoning-Subspace Activation, roughly the share of an image-based internal state that points in the same directions used for text chain-of-thought reasoning. This measure correlates with correctness around 0.37 to 0.43, and image-trained reinforcement raises it most. Swapping visual representations from the trained model into the frozen base model also helps most for that version, supporting the claim that visual features are better aligned to trigger reasoning.
Limitations are real. Supervised and preference variants sometimes hurt on harder sets — for Qwen3-VL-8B they improve MathVerse but degrade MathVision and EMMA — so the method’s success depends heavily on the exploratory objective. Training data is still textbook math rendered as clean images, which underrepresents messy real-world ambiguity, occlusion, and domain coverage. The internal subspace analysis is correlational and does not prove causation for all reasoning. Absolute performance on hard sets like EMMA and ChartQAPro remains low even after gains.
Why this might matter to me
For your interests in spatial reasoning, world models, and 3D generation, the central lesson is that perception and reasoning can be decoupled: a model can correctly list objects and still fail to chain spatial relations.
The case studies make this concrete. One model correctly reads a square of side 4 cm and a triangle of equal area, correctly computes a local triangle height of 8, but forgets the square itself sits 4 cm above the base line, so it answers 8 instead of 12. Only the image-conditioned reinforcement version adds the offset. That is exactly the local-to-global failure you see in text-to-CAD and parametric geometry, where a model names parts correctly but violates a global constraint like concentricity or c1 + c2 = c3.
For world models and scene representation, two practical ideas transfer. First, you could mine your own counterfactuals without manual 3D labels: render a parametric sketch or scene description to an image or view, keep text-success and image-failure pairs, and reinforce reasoning from the render. Second, the activation finding suggests monitoring whether visual tokens actually increase attention and project into text reasoning directions across layers, rather than assuming a good encoder is enough. If you build text-to-CAD pipelines, explicitly training the visual-to-reasoning pathway for offsets, nesting like counting concentric ear circles, and height accumulation may matter more than higher-resolution vision alone.
Terms worth knowing
Multimodal Large Language Model: A system that combines a vision encoder for images with a language backbone for text reasoning, so it can take both pictures and words as input.
Reasoning-access gap: The performance drop from text to semantically identical image that remains even when reading is correct, interpreted as failure to route seen evidence into existing reasoning ability.
Modality counterfactual: A pair made by rendering the same content in a different input form, here text versus a picture of that text, to isolate the effect of modality.
OCR accuracy: How faithfully the model transcribes characters from an image, used here as a perception control separate from reasoning.
Extracted-text control: Reasoning over the model’s own transcription of the image, to test whether restoring text form restores performance.
TIGER: Text-to-Image Gap-targeted training for Enhanced Reasoning, the proposed pipeline of mining text-success and image-failure pairs then training on image inputs.
Supervised Fine-Tuning: Learning by imitating a fixed correct answer token by token.
Direct Preference Optimization: Learning by preferring a good answer over a bad answer for the same prompt, without exploring new answers during training.
Group Relative Policy Optimization: Learning by sampling a group of attempts for one prompt and reinforcing the relatively better ones according to a verifiable reward for correct final answer and format.
Reinforcement Learning with Verifiable Reward: Reinforcement learning where correctness can be checked automatically, as in math with a known answer, rather than by human judgment.
Reasoning-Subspace Activation: The fraction of an image-conditioned internal representation that lies in directions associated with text chain-of-thought reasoning, used as a proxy for whether vision is engaging reasoning.
Activation patching: An intervention test where internal states from a trained model are inserted into a base model to see if they alone improve accuracy.