RightSignal · Sunday, November 1

Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs

Donghyeon Cho, Joowon Kim, Seungho Shin, Eunho Yang · Submitted 27 Oct 2026 · 100 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

Diffusion vision-language models reread the whole image at every step, even though image tokens make up most of their input. LOWN first predicts without the image and only looks at the image when that prediction disagrees with the last one made with it.

Key finding: It keeps 99.6% of LLaDA-V's score across ten benchmarks while using 35.3% of the decoding compute, with no training.

Why it matters: This is a cheap, training-free way to make image-reading models much less costly to run.

Read on arXiv

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Randall Balestriero, Yann LeCun · Submitted 29 Oct 2026 · 58 upvotes yesterday

Preprint · not peer reviewedTrending

The authors prove that models of a type called JEPA, trained separately on different kinds of data, learn the same picture of the world up to a single rotation, as long as their training pairs are two views of the same scene. They argue that next-word prediction is a poor version of this setup.

Key finding: Good views needed 42 times fewer training samples than next-token prediction to reach the same alignment. A 109M-parameter text model aligned with image encoders better than language models of up to 72B parameters.

Why it matters: It suggests that how training examples are paired may matter more than sheer scale.

Read on arXiv

Youtu-Parsing-Omni: One Encoder, One Schema, All Modalities

Yunhang Shen, Zhangquan Chen, Xiaotian Li, Zhongpeng Cai, Qing Zhou, Zekai Zhang · Submitted 29 Oct 2026 · 72 upvotes yesterday

Preprint · not peer reviewedCode releasedTrending

A single 5B-parameter model turns documents, images, audio and video into structured records marked with where and when things appear. It uses one encoder for every kind of input and one output format for all ten parsing tasks.

Key finding: It scores 96.96 overall on OmniDocBench v1.6 and leads open-weight models on OmniParsingBench, with its biggest advantage on audio and video.

Why it matters: AI agents that need to read and cite evidence from many kinds of files could use one parser instead of several.

Read on arXiv

Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers

Jiongfan Yang, Weichen Fan, Haiwen Diao, Zhongang Cai, Ziwei Liu · Submitted 29 Oct 2026 · 36 upvotes yesterday

Preprint · not peer reviewedTrending

SpecSlot is an image generator that splits each image patch into slots by frequency bands from the discrete cosine transform. Only the coarse, low-frequency slot talks to the whole image; fine detail is handled locally.

Key finding: On 512-pixel ImageNet it improves image quality (FID) by 2.84 over a comparable baseline while using 3.9% less compute.

Why it matters: Adding fine detail no longer makes the expensive whole-image attention step grow, which makes high-resolution generation cheaper.

Read on arXiv

Systems

Glance: The Generator Is Its Own Reranker

Adilet Metinov, Chingiz Arziev · Submitted 29 Oct 2026 · 11 upvotes yesterday

Preprint · not peer reviewedTrending

Search-backed chatbots usually run one model to rank retrieved passages and a second to write the answer. Glance adds a small adapter so the answering model ranks the passages itself, reusing work it has already done on the shared context.

Key finding: With one 9B model it beat the bge-reranker-v2-m3 ranker on SciFact and NFCorpus, using about 0.1 GiB of extra memory per skill.

Why it matters: Running one model instead of two could cut memory use and cost for retrieval-augmented systems.

Read on arXiv

Your RightSignal

Topics in News

Look Only When Needed: Certificate-Guided Visual Routing for Efficient Diffusion VLMs

Proving and Improving the Platonic Representation Hypothesis with JEPAs

Youtu-Parsing-Omni: One Encoder, One Schema, All Modalities

Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers

Glance: The Generator Is Its Own Reranker

That’s every paper.