Two papers from HuggingFace Daily Papers converge on an operational point: evaluating vision-language models while staying in text doesn’t hold up. One asks the model to draw the answer, the other measures movement decomposition in videos.

What happened. “Show, Don’t Tell” introduces ProVisE, a framework that lets image generation models answer spatial reasoning questions directly in pixels, then converts the images into the structured predictions that existing benchmarks require. On 470 samples and 14 subtasks, systems that respond visually solve 37% of spatial cases that the reference text model gets wrong. The second paper, by Lukas Knobel with Andrew Zisserman, proposes a self-supervised method to separate camera movement from object movement in videos: two dynamics that until now have been tangled in frames and difficult to annotate by hand.

Why it matters to you. If you use agents that process visual material (scanned documents, product photos, inspection videos), the benchmark score can hide a gap: the model knows how to describe what it sees in words, but can’t position it in space. When the task requires pointing out where a defect is or tracing a path, the text interface becomes the bottleneck. The two papers shift the question from “can the model answer?” to “can the model show?”.

If you want to try it. ProVisE is public and agnostic to benchmarks: the repo and project page are the starting point for seeing how to move from text output to visual output on your own spatial cases.

In detail

Benchmarks for vision-language models work like this: you show the model an image and ask it a question. The answer is text. If the question is “how many red objects are there?”, text is enough. If the question is “draw the path from the door to the table”, forcing the model to translate a spatial position into numerical coordinates or words is a forced interface.

The interface problem. “Show, Don’t Tell” starts here. Many spatial insights are natural to show (point to a location, trace an area, draw a path) and unnatural to write (decimal coordinates, rough verbal descriptions). The paper calls this phenomenon answer-interface mismatch: the model might know where something is, but the text channel forces it into a translation that introduces error.

ProVisE solves it this way: let an image generation model answer by drawing, then convert the visual result into the structured format the benchmark requires. If the benchmark wants a point’s coordinates, ProVisE extracts them from the pixel the model colored. If it wants an area, it derives it from the drawn region.

The main result, reported in the paper’s abstract (HuggingFace, July 24, 2026): on 470 samples across 14 subtasks of SpatialGen-Bench and six external benchmarks, 31 systems evaluated. The two channels, visual and text, are complementary. The visual one wins where spatial state expresses directly in pixels. The text one stays stronger on compositional reasoning, where you need to combine multiple logical rules. GPT Image 2 solves 37% of spatial cases that GPT-5.4, the text model, gets wrong.

The second paper: separating what moves. The work by Knobel, Zisserman, and Asano tackles a different but connected problem. When you watch a video, each frame changes relative to the previous one. That change has two causes: the camera moved (you shifted with the phone) or objects moved (a person entered the shot). In natural videos the two causes are tangled: an object that seems to scroll right might be still while the camera pans left.

Separating the two factors helps build robust movement representations, where object dynamics are decoupled from the camera effect. Until now separation required manual supervision: someone had to label what was camera and what was object. The paper proposes a self-supervised method that learns the decomposition directly from videos, without labels.

What changes for builders. If you’re assembling an agent that processes visual material, the operational angle is clear: your system’s output channel must match the type of answer you need. A model that knows how to describe where a defect is in an image in words isn’t the same as one that knows how to circle it. A model that recognizes “there is movement” in a video isn’t the same as one that can say “the object moves, the camera is still”.

The limits. ProVisE evaluates 31 systems on 470 samples: it’s a test bed, not mass adoption. The complementarity between visual and text means neither channel dominates: you need to know which to use for which subtask. The paper on video decomposition is a representation result, not a ready-to-use tool: it shows that separation is learnable, but doesn’t release it as a module you can plug into an existing pipeline.

To build your own test bed instead of relying on public benchmarks, the playbook on comparing two models and the lesson on how to evaluate your agent without benchmarks start from the same principle: measure your case, not the average.