Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-08-30 · source-backed
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)
Each link below shares sources, entities, or timing with this story.
Niclas Lietzow, Danielle Bitterman, and Carsten Eickhoff probe what happens when a vision-language model's eyes disagree with its memorized world knowledge, identifying a "vision-default, prior-override" causal mechanism. This is directly useful for debugging the maddening cla...
Researchers systematically evaluate whether Mamba-class state space models can replace ViT encoders in large VLMs, finding competitive performance with linear-time processing versus ViT's quadratic attention. Meaningful memory savings on high-resolution or long-context vision...
A 1.5B distilled model trained with GRPO chooses NoThink, Short, or Long at response start, using a shaped reward that makes each mode pay off at a different length plus hard per-mode token caps. Accuracy held at 0.782 against 0.796 baseline while mean length fell from 4,796 t...
Earlier methodologies misclassified standard-library modules as hallucinations (arXiv 2608.22652). Testing seven inference-time defenses across eight models in five families and four languages, RAG reduced the hallucination rate in 18 of 32 model-language configurations, while...
The authors formalize Dense Same-Class Attribute Misbinding and built InstaBind-Lite to measure it: 524 images, 529 groups of 3-6 same-class entities, 9,580 deterministically evaluated questions with source-instance annotations that separate copying from hallucination (arXiv 2...
This paper argues the missing axis is whether a summary satisfies a specific reader, and that a persona is a more practical signal than a query since users rarely state everything relevant (arXiv 2608.14457). A biomedical researcher and a family doctor reading the same vaccine...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.