Does the model actually need to see it?

It is tempting to think a good description can stand in for the screenshot. Paste the error, summarize the chart, transcribe the table, then ask the model to reason over text. For tasks that depend on layout, spatial relationships, or visible state, that substitution can remove the very evidence the question is about.

Take the image away

He et al. ran GPT-4o on ChartQA twice: once with the chart and once as a blind baseline, answering without the image¹¹Distill Visual Chart Reasoning Ability from LLMs to MLLMs (ReachQA) He et al., 2024. The difference was large:

The blind score is not zero because some answers can be guessed from priors or format. What the comparison establishes is narrower and more useful: information omitted from the input cannot be recovered by better phrasing alone.

Better models do not erase the input question

A 2025 evaluation tested GPT-5 on 107 difficult questions from the CHART-6 benchmark. The newer model was more accurate than GPT-4o, but replacing the question with a generated description of the chart produced no reliable gain over asking with the image and sometimes performed worse⁴⁴GPT-5 Model Corrected GPT-4V's Chart Reading Errors, Not Prompting Yang & Chen, 2025. The tested scope is small, yet the design lesson is sound: improve the model if you can, but do not remove source evidence the task depends on.

A screenshot and a transcript carry different things

A screenshot preserves layout, adjacency, emphasis, and shape. A transcript preserves words. In the DocVLM evaluation, adding a learned OCR modality helped models recover text under a constrained image budget, while full visual context still mattered²²DocVLM: Make Your VLM an Efficient Reader Shpigel Nacson et al. (AWS AI Labs), 2024. That is evidence for complementarity, not proof that any one payload format is universally best.

Seeing is not the same as understanding

BLINK makes the opposite boundary visible. On perception tasks that people solve at roughly 96%, the best models in that study were near 50%³³BLINK: Multimodal Large Language Models Can See but Not Perceive Fu et al., 2024. Supplying an image preserves evidence; it does not guarantee the model will interpret that evidence correctly.

The Noru decision

Noru removes the manual screenshot handoff. When a compatible coding agent needs current visual evidence, Noru can provide a fresh screen image together with recognized on-screen text. The agent still has to reason over it, and the model can still be wrong.

When a question depends on visible layout, labels, or relationships, a fresh screen image preserves evidence that a description may omit. These studies motivate Noru's design; they do not measure Noru's effect on task completion.