Making Wordless Reasoning Look at the Image ── ReaLVR measures and corrects latent visual reasoning's insensitivity to visual evidence

1. Reasoning that is never put into words cannot be checked for what it looked at

Multimodal large language models (MLLMs), which handle both images and text, often answer a question about a figure or photo by first writing out their intermediate thinking, a so-called chain of thought. Latent visual reasoning (LVR) takes a different route: the intermediate computation runs in continuous "latent tokens" instead of words. The hope is that this saves the cost of writing everything out and lets the model keep visual information that is hard to express in language.

There is a price. A written chain of thought can be read and inspected for where the model looked and how it reasoned. Latent tokens cannot be read, which makes it hard to confirm from outside whether the model is actually reasoning from evidence in the image, and hard to supervise what those tokens should learn. The paper starts by measuring exactly that difficulty. The primary source is a preprint posted on arXiv and has not been peer reviewed.^[1]

2. The problem found, and the core of the proposal

The authors first analyze how latent tokens behave and identify what they call a latent evidence-credit gap: when the image is altered in a way that changes the correct answer, the latent tokens respond only weakly. The part of the model that is supposed to be reasoning about the image is insensitive to changes in the image.

They hypothesize that the cause is the absence of explicit supervision during GRPO training. GRPO is a reinforcement learning method that compares several sampled responses and rewards them on the quality of their final answers.^[5] In the authors' view, a reward on the final answer alone says too little about which visual evidence to preserve, or how credit should be assigned across the latent tokens.

Their answer is ReaLVR, which applies visual-evidence supervision to the model's own free-running latent trajectories. It relies on two contrasts. Comparing the correct answer with wrong answers the model itself generated determines where stronger supervision is needed. Comparing relevant visual evidence with mismatched evidence specifies what should be preserved.

3. What the abstract reports

Across three model families, ReaLVR is reported to consistently outperform the LVR baselines it was evaluated against. On Qwen2.5-VL-7B it reached a five-task average of 63.7%, the highest among the methods compared.

The authors also claim to be the first to scale visual reasoning in latent space, reporting that the improvements continued at frontier scales up to 235B. The abstract does not state the size of those improvements. The "first" framing is likewise the authors' own claim, not something independently confirmed.

Further analyses point to more question-sensitive latent-token positions, stronger alignment with relevant image regions, and greater fixed-context dependence on the most attended latent tokens. All three are offered as support for the reading that the latent tokens now make more use of visual evidence.

4. A concrete case: asking a model to read a chart

Suppose an MLLM is given a clinical trial figure from product materials, for example a chart showing how two groups change over time, and is asked which group comes out ahead. A latent reasoning model does not write out its thinking, so even when its answer is right there is no way to tell whether it came from reading the chart.

The first model may be answering from the wording of the question or from patterns learned in training rather than from the chart. The evidence-credit gap described in the paper can be understood as this same insensitivity, observed at the level of individual latent tokens. ReaLVR tries to reduce it by supervising the places that ought to respond to such edits.

5. What is not new, and what the abstract leaves open

Editing the input and checking whether the output changes has long been used to probe what a model's decisions rest on. Contrasting correct and incorrect answers to sharpen a training signal is also a familiar idea. The paper's contribution is better read as applying these to the model's own generated latent trajectories, and designing separately for where to supervise and what to preserve.

Several limits stand out. First, the abstract does not explain how the gap was measured: which edits were used, and how response strength was defined. Second, it gives no figures for the baselines that the 63.7% should be compared against, and it does not name the five tasks. How large the margin is cannot be judged without the full paper.

Third, even with these gains, latent tokens do not become readable. Stronger alignment with image regions suggests that the model's internals draw more on visual evidence, but it does not mean a person can now inspect the reasoning. Alignment is also a correlational measure, not evidence of cause. Fourth, the explanation that GRPO's lack of supervision is the cause is labeled by the authors themselves as a hypothesis.

Fifth, the implementation repository is public, but at the time of checking its description listed trained models as "coming soon".^[3] For now, an independent party reproducing the results under the same conditions would have to train from scratch.

6. Why this topic is forming a cluster now

In this collection cycle, one more paper from a different group landed in the same cluster, giving a cluster size of 2. Scaffolding Minds, by its title, optimizes latent visual target representations to improve multimodal reasoning; it works on what the latent tokens should represent from the representation side.^[2] Where ReaLVR approaches the question by supervising latent reasoning with visual evidence, Scaffolding Minds approaches it by shaping the representation the latent tokens should aim for.

What the two share is the concern that latent reasoning trained only on answer correctness leaves it unclear what the model is looking at. On the paper-sharing site the paper has 98 upvotes, and the implementation repository has 7 stars.^[4] These are signs of attention. They do not show that the claims are correct or important.

7. Where this connects to pharmaceutical and regulatory work

When generative AI reads charts and tables during promotional material review or medical information responses, getting the answer right is not enough. It must be possible to show afterward which part of the source the answer rests on. A model that writes out its reasoning at least leaves an explanation behind. Latent reasoning gives up that written trail in exchange for efficiency and performance.

Within that direction, this paper shows a way to make the model actually use the visual evidence. But using evidence and being inspectable by a person are separate things. For regulated work, it is not enough to wait for latent reasoning to improve; there should also be a separate mechanism that makes the model point to the passages it relied on, and a step in which a person checks the primary source.

The most directly usable part is the check shown in the example: edit the content of a chart or table so that the correct answer changes, and see whether the model's answer follows. That test works regardless of model type and can be built into pre-deployment validation. Given that this research has not yet been peer reviewed, the paper is best read less as a method to adopt and more as material for deciding what needs to be checked.