Escape Meridian

Vohra, Siddharth, Xu, Min · Nov 17, 2026

Vohra and Xu ask when vision-language models need to zoom

Markdown · JSON · Original source

In their paper, Siddharth Vohra and Min Xu argue that image handling before a vision-language model sees a picture can erase small details—and that zooming can recover much of what is lost. A car 20 pixels wide in an 8000 by 6000 drone photo reaches about 6 pixels wide, they report, while 's default rejects the photo. Across tests of seven models from four vendors, the smallest text read half the grew from 6 pixels or less on a small image to a predicted 10 to 13 pixels on a large one at default settings. The authors compare splitting a photo into tiles with a two-step zoom, where the model first identifies the region it needs. Tiling raised accuracy for Claude and Gemini; for GPT-6.1 Sol, avoiding rejections raised it from 50.7% to 83.7%. The two-step approach matched or beat tiles on Gemini and came within 2.5 points on Claude, using about 40% of their image tokens. The findings make the practical question not just whether to zoom, but when and how. The paper also highlights a risk beyond missed details: when detail is lost, Gemini gives a confident wrong answer more often than Claude, though both often say they cannot tell. The authors report that tiling only after Claude replies unsure comes close to always tiling while using fewer image tokens. They turn their findings into five rules for when an agent should zoom.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

arXiv cs.AIArticle · arxiv.orgOpen original source

Original source

cs.AI · Article

Source details

Original source date: Nov 17, 2026 (2026-11-17).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-18T17:16:44.070Z. First observed: 2026-11-18T05:07:18.650Z. Retrieved: 2026-11-18T17:06:26.064Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

This reading is based on the page's abstract, not the full paper or its underlying data. The abstract describes tests on seven models from four vendors; it does not establish how the results transfer to other models, images, or uses.

Open question

How well do the reported zoom strategies and model differences carry over to images and uses beyond the tests described in the abstract?

Source-stated

Vohra and Xu report that image resizing or rejection can make small details harder for vision-language models to read.

Online vision-language models often miss small details in large images, because the service shrinks the image, or rejects it, before the model ever sees it.

Source-stated

The authors tested seven models from four vendors using printed-code reading and questions about small objects in real photos.

We measure how much detail survives for seven models from four vendors, using an eye exam of printed codes and 601 questions about small objects in real photos.

Source-stated

Tiling improved accuracy for Claude and Gemini, and avoiding rejections improved GPT-6.1 Sol's accuracy.

Splitting each photo into up to nine tiles raises accuracy by 11.1 to 14.5 percentage points for Claude and Gemini, and by avoiding rejections it lifts GPT-6.1 Sol from 50.7% to 83.7%.

Source-stated

A two-step zoom matched or beat tiling on Gemini and came close on Claude while using fewer image tokens.

A two-step zoom, in which the model first names the region it needs, matches or beats tiles on Gemini and comes within 2.5 points of them on Claude, with about 40% of their image tokens.

Source-stated

When detail is lost, the authors report that Gemini gives confident wrong answers more often than Claude.

Gemini gives a confident wrong answer in 43% to 48% of these cases, Claude in only 5% to 9%.

Source-stated

The abstract reports that tiling only after an unsure Claude reply came close to always tiling while using fewer image tokens.

tiling only after an unsure reply comes within 1.4 points of always tiling with 62% to 70% of the image tokens.

Source and contributor context

Vohra, Siddharth, Xu, Min · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Claude

Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.

OpenAI

OpenAI develops AI models and products and publishes research, release announcements and evaluations.

TIME

TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.

Editorial provenance

Version 1. Publication ID: 7aa40093-41ff-42ec-8288-3ba48f9c902f.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-18T17:16:44.070Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 68f9b6d6dea070d5faaae6be0055979ec4d00b299786c122b2551240dfb796c6. Source text hash: b0c0e43eb0ccea1189fde0c9eff057817ccb611025a747438be8a0044634d768.

Revision and source-check history

Version 1 · 2026-11-18T17:16:44.070Z

Vohra and Xu ask when vision-language models need to zoom

In their paper, Siddharth Vohra and Min Xu argue that image handling before a vision-language model sees a picture can erase small details—and that zooming can recover much of what is lost. A car 20 pixels wide in an 8000 by 6000 drone photo reaches about 6 pixels wide, they report, while 's default rejects the photo. Across tests of seven models from four vendors, the smallest text read half the grew from 6 pixels or less on a small image to a predicted 10 to 13 pixels on a large one at default settings. The authors compare splitting a photo into tiles with a two-step zoom, where the model first identifies the region it needs. Tiling raised accuracy for Claude and Gemini; for GPT-6.1 Sol, avoiding rejections raised it from 50.7% to 83.7%. The two-step approach matched or beat tiles on Gemini and came within 2.5 points on Claude, using about 40% of their image tokens. The findings make the practical question not just whether to zoom, but when and how. The paper also highlights a risk beyond missed details: when detail is lost, Gemini gives a confident wrong answer more often than Claude, though both often say they cannot tell. The authors report that tiling only after Claude replies unsure comes close to always tiling while using fewer image tokens. They turn their findings into five rules for when an agent should zoom.

Original source

cs.AI · Article

Source details

Original source date: Nov 17, 2026 (2026-11-17).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-18T17:16:44.070Z. First observed: 2026-11-18T05:07:18.650Z. Retrieved: 2026-11-18T17:06:26.064Z.

Limits

This reading is based on the page's abstract, not the full paper or its underlying data. The abstract describes tests on seven models from four vendors; it does not establish how the results transfer to other models, images, or uses.

Open question

How well do the reported zoom strategies and model differences carry over to images and uses beyond the tests described in the abstract?

Source-stated

Vohra and Xu report that image resizing or rejection can make small details harder for vision-language models to read.

Online vision-language models often miss small details in large images, because the service shrinks the image, or rejects it, before the model ever sees it.

Source-stated

The authors tested seven models from four vendors using printed-code reading and questions about small objects in real photos.

We measure how much detail survives for seven models from four vendors, using an eye exam of printed codes and 601 questions about small objects in real photos.

Source-stated

Tiling improved accuracy for Claude and Gemini, and avoiding rejections improved GPT-6.1 Sol's accuracy.

Splitting each photo into up to nine tiles raises accuracy by 11.1 to 14.5 percentage points for Claude and Gemini, and by avoiding rejections it lifts GPT-6.1 Sol from 50.7% to 83.7%.

Source-stated

A two-step zoom matched or beat tiling on Gemini and came close on Claude while using fewer image tokens.

A two-step zoom, in which the model first names the region it needs, matches or beats tiles on Gemini and comes within 2.5 points of them on Claude, with about 40% of their image tokens.

Source-stated

When detail is lost, the authors report that Gemini gives confident wrong answers more often than Claude.

Gemini gives a confident wrong answer in 43% to 48% of these cases, Claude in only 5% to 9%.

Source-stated

The abstract reports that tiling only after an unsure Claude reply came close to always tiling while using fewer image tokens.

tiling only after an unsure reply comes within 1.4 points of always tiling with 62% to 70% of the image tokens.

Source and contributor context

Vohra, Siddharth, Xu, Min · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Claude

Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.

OpenAI

OpenAI develops AI models and products and publishes research, release announcements and evaluations.

TIME

TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.

Editorial provenance

Version 1. Publication ID: 7aa40093-41ff-42ec-8288-3ba48f9c902f.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: passed by anthropic/claude-opus-5.5 on 2026-11-18T17:16:44.070Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 68f9b6d6dea070d5faaae6be0055979ec4d00b299786c122b2551240dfb796c6. Source text hash: b0c0e43eb0ccea1189fde0c9eff057817ccb611025a747438be8a0044634d768.

passed · 2026-11-18T17:16:44.070Z · anthropic/claude-opus-5.5

Publication: 7aa40093-41ff-42ec-8288-3ba48f9c902f

attribution — pass: The paper and its findings are correctly attributed to Siddharth Vohra and Min Xu, the listed authors. Results are framed as what the authors report or argue, not as established fact.

Authors:Siddharth Vohra, Min Xu

summary — pass: The reading leads with the actors and the core argument: services shrink or reject large images before the model sees them, and zooming recovers detail. It explains tiling and two-step zoom in plain terms, and keeps the key numbers, the comparison of failure modes and the five rules. Leaving out the finding that a size-based pre-call check misses most of the gain is acceptable. The title and question are neutral. Image tokens could use a tooltip, but the term does not block understanding.

Zooming wins much of the detail back.

date — pass: evidence.originalPublished is 2026-11-17. This matches the arXiv v1 submission date, not the later feed date.

[Submitted on 17 Nov 2026]

support — pass: Every claim matches the abstract, including the figures: 50.7% to 83.7%, within 2.5 points, about 40% of tokens, and 62% to 70% of tokens. The Gemini vs Claude contrast in confident wrong answers is supported. The line that the practical question is when and how to zoom follows from the authors' five rules. The limitations correctly note that the reading rests on the abstract only.

Gemini gives a confident wrong answer in 43% to 48% of these cases, Claude in only 5% to 9%.

media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the original source.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.