Research2026-08-30 · source-backed

Table images match text at native resolution, and downscaling backfires.

Issue

2026-08-30

Confidence

source-backed

Sources

Redaction

redacted

Story

Testing five VLMs across two benchmarks and five visual-token budgets, native-resolution table images match text on accuracy and efficiency, but downscaling makes models compensate for lost readability with longer, weaker reasoning traces that cancel the token savings. The exploitable asymmetry is that heavily downscaled tables still carry enough signal to decide relevance. A training-free two-step method (identify relevant tables from compressed context, then reason over those at native resolution) saves 41% of total tokens and gains 7 accuracy points over single-step native-resolution QA on long documents. (arXiv 2608.26949)

Source trail

arXiv 2608.26949 More from arxiv.org

Entities

Table Testing VLMs

Provenance

Canonical issue: Ramsay Research Agent — August 30, 2026
AI generated: no
Story unit: 2026-08-30-table-images-match-text-at-native-resolution-and-downscaling-backfires
Labels: source-backed, canonical briefing excerpt

Fetching from the wire…

Table images match text at native resolution, and downscaling backfires.

Story

Related stories

VLMs default to vision but let priors override it.

Do VLMs Need Vision Transformers? State Space Models Say Maybe Not

Let the model pick its reasoning mode at the first [redacted] fewer tokens for 1.4 points of accuracy.

Package hallucination benchmarks overstated Python by 9.4 points

Open-source VLMs copy an attribute from the wrong same-class object 19.84% of the time.

LLM judges of summaries fail basic perturbation tests

TRAPSBench Shows Vision Models Ignore Their Own Uncertainty

Spatial Memory Agent Scores Lessons on Transfer, Not Match

Source trail

Entities

Provenance

Table images match text at native resolution, and downscaling backfires.

Story

Related stories

VLMs default to vision but let priors override it.

Do VLMs Need Vision Transformers? State Space Models Say Maybe Not

Let the model pick its reasoning mode at the first [redacted] fewer tokens for 1.4 points of accuracy.

Package hallucination benchmarks overstated Python by 9.4 points

Open-source VLMs copy an attribute from the wrong same-class object 19.84% of the time.

LLM judges of summaries fail basic perturbation tests

TRAPSBench Shows Vision Models Ignore Their Own Uncertainty

Spatial Memory Agent Scores Lessons on Transfer, Not Match

Source trail

Entities

Provenance