When generation is the reasoning: notes on VBVR-Pro

The premise: generation as the substrate, not the output

Most of what we call "visual reasoning" today is really language reasoning with pictures attached. A model looks at an image, converts what it sees into tokens, and reasons in text. The image is an input to be understood, or, at best, an output to be rendered at the end of a pipeline.

VBVR-Pro, from a 52-author team spanning labs and universities including groups working on video generation, embodied cognition, and multimodal evaluation, argues for something different: treat the visual generation process itself as the medium of reasoning. Instead of "think in text, then draw the answer," the model reasons by generating — producing a sequence of images, video frames, or interleaved image-text steps where each step is a state in a search or inference process, not a final illustration. Think of a video model imagining an object rotating to check whether it fits through a gap, rather than describing the geometry in words.

That's a compelling reframing, but it has been hard to study rigorously. The paper's diagnosis of why is the more useful contribution: three infrastructure gaps have been quietly blocking progress.

Three gaps, one closed loop

No scalable training tasks. Visual reasoning benchmarks tend to be small, hand-curated, and static — fine for evaluation, bad for reinforcement learning, which needs a large, controllable task distribution to train against.

No reliable feedback signal. The default way to score a generated image or video against a task is to hand it to another multimodal LLM and ask if it looks right. VBVR-Pro's authors ran a systematic study of leading MLLMs used this way and report recurring, characteristic failure modes in that judge-based paradigm — the kind of noisy reward signal that quietly caps how far RL can push a model.

No controlled comparisons across substrates. Image generation, video generation, and interleaved generation are usually evaluated in separate papers with separate setups, so claims like "video reasoning is more powerful" rarely get an apples-to-apples test.

VBVR-Pro's answer is to build one closed loop that addresses all three at once: a procedurally generated task library, deterministic reward scorers instead of VLM judges, and a shared harness across more than 30 generative models spanning image, video, and interleaved modalities.

What's actually in the suite

Task scaling. 300 procedurally generated tasks form a controlled space rather than a fixed, hand-picked set — procedural generation means the task distribution can be scaled or reshaped without new human annotation for every item. The authors report that models trained on VBVR-Pro transfer to seven external visual reasoning benchmarks, naming RISE-Video, MME-CoF-Pro, and BabyVision among them. That transfer claim is the headline result for anyone deciding whether this is worth training against, and it's the one place I'd want the actual numbers before treating it as settled — the abstract asserts "strong transfer" without reporting the scores.

Verifiable rewards. This is the part I find most defensible on its face, independent of the missing numbers. Grounding reward in deterministic, task-specific rules — did the object end up in the right place, does the trajectory satisfy the constraint — sidesteps the well-known brittleness of VLM-as-a-judge scoring, where a judge model's own blind spots become the reward function's blind spots. The paper reports fine-grained alignment with human judgments for these rule-based scorers, and downstream, better post-RL performance than judge-based rewards produce. Deterministic reward is also just more practical at RL scale: it's cheap, doesn't drift, and doesn't require running a second large model as part of every reward computation.

Mechanism study. With a shared task suite and a shared scorer, the authors run a controlled comparison across generation modalities. Two findings stand out. Video generation is strongest on tasks that require persistent spatiotemporal state tracking — the kind of problem where you need to remember where something was three steps ago. Interleaved generation (alternating image and text steps) comes out as a compute-efficient alternative, presumably because it doesn't pay the cost of generating dense video frames for state that could be captured more sparsely. The paper also reports ablation and probing evidence for "vision-native trajectories" that matter for reasoning quality — internal generative paths that aren't simply reducible to language-mediated planning. That's the most interesting mechanistic claim in the abstract, and also the least specified; I'd read the paper itself before leaning on it.

The honest caveat

The abstract is unusually confident in its framing and unusually thin on numbers. "Strong transfer," "stronger post-RL performance," "fine-grained alignment" — these are all claims I'd want to see quantified before citing them in a design decision. That's not a knock on the paper's value; the infrastructure — the task generator, the deterministic scorers, the multi-substrate harness — is real and independently useful even if some headline claims land softer than the abstract suggests. Open data, models, scorers, and code lower the cost of checking for yourself, which is the right way to treat any single benchmark paper's claims.

Why this matters if you're building agents

"Reasoning-via-generation" is a small but growing niche, and it's been starved of the same thing every RL-adjacent subfield needs early: a shared, scalable, verifiable task suite that lets different approaches actually be compared instead of eyeballed. If you're evaluating whether an image or video model can serve as more than an illustrator — whether it can function as a reasoning engine over visual state — VBVR-Pro is now the natural place to run that comparison, and the choice to ground reward in deterministic rules rather than a VLM judge is the design decision worth borrowing regardless of whether you use this exact suite.