Preprint

V2T

Distilling visual reasoning into text space

Train with images in the chain of thought. Reason without them at inference.

Wenhan Yang1,  Nilay Naharas1,  Ali Payani2,  Baharan Mirzasoleiman1

1University of California, Los Angeles  2Cisco Research

Paper Code coming soon BibTeX

More accurate, for a fraction of the compute

Average score on a broad visual suite against training GPU-hours or inference cost. V2T sits at the top left: highest score at the lowest cost.

Average over HR-Bench 4K/8K, V*Bench, BLINK and POPE, all 7B models (Table 2 of the paper). Hover or tap a point for its name. Costs marked ∗ in the paper are estimated from reported configurations.

63.4

average on the broad visual suite, the best of all methods compared, from tool-based to latent visual reasoning

42×

fewer GPU-hours than Vision-R1 (80 vs 3,392), and 1× inference cost with no intermediate images

73.3

on MathVista, the highest among the 7B reasoning models we compare, after a short RL run

Overview

Many visual problems need more than the picture you are given: an auxiliary line in a geometry diagram, a zoomed-in crop, the next board state. Recent models generate those intermediate images, or latent visual tokens, inside their chain of thought.

That is slow, the generated visuals can be wrong, and errors carry forward. We also find that interleaved images increasingly get in the way of the text reasoning itself. V2T keeps the benefit and drops the cost: a teacher learns from interleaved image-and-text reasoning, and a student absorbs that skill while reasoning in text alone. At inference it is as cheap as any text-only model, and it beats its own teacher.

A geometry problem solved step by step. Each text step is paired with an image: first the incircle with radii drawn to the tangency points, then a line from B to the center O.

Scroll sideways to see the whole chain →

An interleaved chain of thought (Figure 1 of the paper): each reasoning step comes with an image, here the auxiliary constructions for a geometry problem. V2T learns from data like this, but never has to draw these images itself.

But interleaved images get in the way of the text

Longer reasoning, more interference

Extra loss from adding interleaved images, by how far into the chain of thought.

Even blank content hurts

Replacing half of each image with gray makes the gap larger, not smaller.

Text catches up, then wins

Accuracy when continuing from the same ground-truth reasoning steps, with or without the images.

Redrawn from Figure 2 of the paper. Loss gap is the mean loss with interleaved images minus without.

Method

V2T uses interleaved images as training-time supervision only. Two short supervised stages, then an optional short RL stage.

Training diagram. The frozen teacher reads the question with ground-truth reasoning images; the student reads the question image only. Logit KL links their token distributions at every step, and attention KL links the teacher's attention on each reasoning image to the student's attention on the matching region of the original image.

Scroll sideways to see the whole diagram →

V2T training (Figure 3 of the paper). The frozen teacher reads the ground-truth reasoning images; the student sees only the original problem. Logit KL aligns every step, and attention KL (red) moves the teacher's focus on each reasoning image to the matching region of the original image.

Stage 1

Train an interleaved teacher

image text steps paired with reasoning images

Fine-tune on chains of thought where steps come with crops, zooms or drawn constructions. The teacher learns where to look.

Stage 2

Distill into a text-only student

match teacher logits (KL) match teacher attention on the image

The student sees only the problem image. It learns from ground-truth text, the teacher's next-token distribution, and the teacher's attention projected back onto the original image.

Optional

Short GRPO with a grounding reward

r = rtask + η rground overlap of attention with the true region

Distillation already gives a strong start, so a short RL run is enough. When boxes are available, the reward also pays for attending to the right region.

Distillation objectiveL = Ltext + λ LKD + β Lattn  λ = 0.1, β = 0.5

Grounding rewardrground = ½ ( |M̂ ∩ M| / |M̂| + |M̂ ∩ M| / |M| )

The teacher is conditioned on ground-truth reasoning images during distillation, never on generated ones, so errors in generated visuals cannot leak into the student.

Examples

Real questions from the paper. All three models get only the question and the original image. Phrases like “zoomed-in view” come from the teacher's interleaved format: V2T never generates or receives an extra image, it reasons over the original one.

Results

Broad visual suite: high-resolution perception (HR-Bench 4K and 8K), visual search (V*Bench), fine-grained perception (BLINK) and hallucination (POPE). V2T is compared with text-reasoning, latent visual reasoning, unified interleaved and tool-based methods.

V2Tother methods
Text-only SFT attention: scattered over the boardwalk and bushes, weak on the post Interleaved teacher attention: spread over bushes on the left, partly on the post V2T attention: concentrated on the post inside the red box

Attention after the first reasoning step, on a Zebra-CoT visual search problem (Figure 5 of the paper). Red box: ground-truth region.

It looks where it should

Attention IoU with the ground-truth box, over 100 Zebra visual search problems. Given its reasoning images, the teacher shifts attention onto them and away from the problem image. V2T grounds best.

Fewer spatial-reasoning errors

Error counts by category, relative to the interleaved teacher. The biggest drops are in spatial and global context.

Attention distillation matters

Broad-suite average. Distilling attention adds 4.7 points, with +14.9 on HR-Bench 8K and +16.7 on V*Bench; the grounding reward adds 1.4 more.

The student beats its teacher

Broad-suite average for the training-matched baselines, all from the same base model and data.

Full comparison

Best per column in bold, second underlined. Train is total GPU-hours; ∗ marks estimates.

Math reasoning

We also apply V2T to visual math, training Qwen2.5-VL-7B on ViRC, where each reasoning step can come with a crop of the problem figure. Because each crop maps back to a box in the original image, the teacher's attention on a crop can be transferred to that box.

Interleaved teacherFull fine-tuning on 99,833 ViRC chains of thought with crops inserted mid-reasoning, 2 epochs, vision tower and projector frozen.
Text-only studentStarts from the base model and sees only the problem image. Loss: cross-entropy + 0.1 × logit KL + 0.5 × attention KL, using 126,740 crop boxes.
Short GRPOA brief RL run from the distilled student. We report a single checkpoint.
V2T + GRPOother 7B models

By benchmark and problem type

MathVista (testmini) overall and by problem type, followed by two separate benchmarks, MathVision (mini) and LogicVista. Best per column in bold, second underlined.

In-domain results on Zebra-CoT

Accuracy on 200 held-out problems per category from the training distribution.

Citation

@article{yang2026distilling,
  title   = {Distilling Visual Reasoning into Text Space},
  author  = {Yang, Wenhan and Naharas, Nilay and Payani, Ali and
             Mirzasoleiman, Baharan},
  journal = {arXiv preprint arXiv:2609.34408},
  year    = {2026}
}