Noru Flow / Research & Notes
On seeing, memory, and control.
Research syntheses and product notes about giving coding agents better evidence, with a clear boundary around every observation.
Follow this publicationN / R
Without image?30.0%
vsWith chart85.7%
GPT-4o · ChartQA · external comparisonFeatured · Research synthesis · Seeing
Does the model actually need to see it?
Published evidenceIn one GPT-4o ChartQA comparison, accuracy was 85.7% with the chart and 30.0% without it. When an answer depends on visual relationships, the image carries evidence a description may omit.
More research
7 publications
01 / Product note · BoundariesA screenshot is evidence, not permissionLetting an agent observe the screen should not quietly grant it the right to act. Noru keeps observation local, visible, and separate from control. 02 / Product note · EvidenceWhy old context cannot answer “now”History can explain what happened before. It cannot prove what is on screen now, and a failed fresh observation must never be replaced with an old one. 03 / Product note · MemoryMemory should return evidence, not confidenceNoru Pro memory is retrieval over local captures, not training. It can suggest a small number of prior observations, preserve their provenance, and return nothing when nothing earns its place. 04 / Noru evaluation · EvidenceWhat 7 of 8 does and does not meanOn one locked synthetic fixture, Noru's selector surfaced relevant prior evidence for 7 of 8 answerable queries and abstained on 2 of 2 no-answer queries. We did not test agent task outcomes. 05 / Research synthesis · SeeingWhy the screenshot needs a text layerAt constrained image budgets, OCR can restore characters the pixels lose while the screenshot preserves layout. The evidence supports a paired payload, not OCR as ground truth. 06 / Research synthesis · SeeingHow much screenshot is enough?In one Qwen2-VL resolution sweep, InfographicVQA rose from 28.9% to 77.3%, while fine-OCR peaked earlier. More pixels can help, but the useful budget depends on the model and task. 07 / Research synthesis · SeeingThe same context, in a better orderIn VisualWebArena's tested GPT-4V baseline, adding screenshots raised success from 7.3% to 15.1%; Set-of-Marks reached 16.4%. The result supports structured visual context, but Set-of-Marks is not a Noru feature.