Visual Reasoning

TL;DR for operators If a multimodal model reaches the correct final answer after misreading part of an image or making an invalid intermediate inference, a binary success reward gives the optimizer little information about what should actually be reinforced. The paper studied here turns those failure classes into separate training signals and, where possible, assigns their reinforcement-learning credit only to the relevant portion of the response. ...

TL;DR for operators A model reads a chart, diagram, or photographed math problem and produces the wrong answer. Treating that event as a generic “reasoning failure” can send engineering effort to the wrong component. The model may have misread a number, attached a label to the wrong object, confused a scale or unit, or reasoned incorrectly after extracting the right facts. ...

TL;DR for operators Longer reasoning is not automatically safer for vision-language systems. This paper finds that a model can preserve the relevant visual information yet perform worse once reasoning shifts attention away from the image region that matters. The paper separates two questions that are often conflated in multimodal evaluation: how much attention goes to the image, and whether that attention is concentrated on evidence relevant to the question. Targeted masking experiments suggest that the heads with concentrated visual focus are functionally important, not merely correlated with success. ...

TL;DR for operators A model can correctly recognize that one object is left of another and still fail when asked to generate that same relationship. More importantly, some visual weaknesses recur in both directions. AMVICC maps visual-language understanding and image generation onto visual concepts derived from the same underlying benchmark. Across the tested systems, weaknesses repeatedly appear in Quantity and Count, Positional and Relational Context, Orientation and Direction, and State and Condition. Text behaves differently: most tested multimodal language models avoid the paper’s Text failure threshold, while all three tested image generators fall below it for explicit generation. ...

A dashboard is rarely confusing because the pixels are ugly. More often, the problem is that the important part is small, crowded, rotated, hidden in a chart corner, split across spatial relations, or buried inside a scene that needs to be mentally transformed before the answer becomes obvious. A human analyst zooms, marks, traces, rearranges, or imagines a new angle. A multimodal model, by contrast, is often asked to stare at the original image and talk harder. ...

A model can generate a pretty sequence of images. Good. So can a slide deck. The harder question is whether those images actually help it think. That is the uncomfortable point behind MentisOculi: Revealing the Limits of Reasoning with Mental Imagery, a new benchmark paper that tests whether frontier multimodal models can do something closer to human mental imagery: form a visual state, keep it stable, transform it step by step, and use the transformed state to decide what to do next.1 Not merely “look at an image and answer a question.” Not “draw a plausible intermediate picture.” Actual visual reasoning, with consequences. ...

A dashboard screenshot is often too little. A video walkthrough is often too much. Somewhere between the two sits a strangely old-fashioned interface: panels, captions, arrows, speech bubbles, and a sequence that tells the machine what happened before what. Yes, comics. That sounds unserious only if we think comics are a decoration layer: something added after the reasoning is complete to make the output friendlier. The paper Thinking with Comics: Enhancing Multimodal Reasoning through Structured Visual Storytelling makes a more interesting claim: comics can act as the reasoning medium itself, not merely the illustration of reasoning after the fact.1 ...

A model can see the image and still miss the point Inspection is a wonderfully cruel test for AI. Show a multimodal model a product photo, a medical scan, a factory defect, a form, or a dashboard screenshot, and the answer may sound calm, fluent, and technically plausible. The model may even imitate the reasoning style of a stronger teacher model. It may describe objects, infer relationships, and produce the correct-looking sentence. ...

Clicks Are Cheap. Wrong Clicks Are Not. Click. That is the unit where many AI agent demos stop being impressive and start becoming expensive. A planning model can write a beautiful instruction sequence: open the settings panel, choose the correct tab, find the export button, confirm the dialog. Lovely. Then the visual grounding model clicks the button two pixels away from the actual target, or chooses the visually similar icon beside it, or mistakes a disabled control for an active one. Suddenly the “agentic workflow” is not a workflow. It is a small robot poking the wrong part of a screen with great confidence. Very modern. Very avoidable, perhaps. ...

A camera on a factory line does not need to write an essay before deciding whether a part is cracked. That sounds obvious. Yet a surprising amount of recent AI architecture quietly assumes the opposite: when vision systems become uncertain, bring in a large language model, ask it to generate richer descriptions, then run the detector again. Sometimes this works. It also turns a detection problem into a small committee meeting, and committee meetings are rarely known for real-time throughput. ...