The same context, in a better order

Two payloads can contain related pixels and words yet be easier or harder to use. Structure affects which evidence a model notices, how it connects text to an interface, and whether important context is buried. The individual techniques need to stay distinct: adding pixels is not the same as labeling elements, and neither is the same as ordering the payload.

Pixels, labels, and order are different interventions

VisualWebArena evaluated a GPT-4V agent on realistic, visually grounded web tasks¹¹VisualWebArena: Evaluating Multimodal Agents on Realistic Visually Grounded Web Tasks Koh et al., 2024:

Adding screenshots more than doubled the tested baseline. Adding Set-of-Marks labels lifted it again. Set-of-Marks is an external technique that overlays visual references on interactive elements; Noru does not add those labels. The absolute scores belong to the tested model, agent, and dataset.

Order is not neutral

Tan et al. found large performance swings when they changed the position of information in multimodal inputs²²Order Matters: Exploring Order Sensitivity in Multimodal Large Language Models Tan et al., 2024. Their experiments concern context position, not OCR reading order. The connection to Noru is a design analogy: evidence should arrive with a deliberate structure, and that structure should be described precisely.

Hard benchmarks define the boundary

MuirBench tests multi-image reasoning; ScreenSpot-Pro tests precise GUI grounding in professional applications³³MuirBench: A Comprehensive Benchmark for Robust Multi-image Understanding Wang et al., 2024⁴⁴ScreenSpot-Pro: GUI Grounding for Professional High-Resolution Computer Use Li et al., 2025. Both show that visual reasoning remains difficult in their tested settings. Neither measures Noru, nor do they show that ordered OCR alone solves the problem.

The Noru decision

Noru currently contributes a fresh image plus recognized text in reading order. It keeps the capture identity, dimensions, and provenance attached so the evidence remains inspectable. It does not add Set-of-Marks labels, and it does not turn an external benchmark into a product-performance claim.

Structure is part of the payload, not decoration. The cited work explains the design direction; it does not measure a Noru performance gain.