← News

Models & toolsSources: arXiv · 19 Nov 2026

V-lens method visualizes hidden concepts inside unified multimodal models

Researchers introduce V-lens, a training-free way to render a model's internal token states as images, plus a tuning method that sharply improves those visualizations.

In plain words

Some AI models both understand and generate images, and their image generators can tap into the internal states formed while reading text. V-lens turns one of those states into a picture, offering a glimpse of the concepts a model links to a question. A tuning step makes these pictures far more accurate.

What was announced

The authors say unified multimodal models, which combine understanding and generation in one architecture, make it possible to inspect reasoning steps by visualizing intermediate representations. V-lens uses the link between a model's image generator and the representations formed while processing text to picture a chosen token's hidden state without training a decoder. Across four diffusion-based and two purely autoregressive unified models, the resulting images showed key intermediate concepts connecting a question to its answer, even when those concepts were never named explicitly.

Controlled tests tracked how this visual content changes with context, token position, and depth. The team then proposed lens tuning, which supervises selected hidden states with pictures through V-lens. On BAGEL, SenseNova-U1, and HunyuanImage-3.0, correct visualizations of pronoun referents rose from 2-11% to 86-94%. The authors also report that visual supervision improved text answers even though answer tokens were not supervised directly, and they released code on GitHub.

What it does not show

  • The source is an abstract, so experimental details, model versions, and evaluation methods cannot be verified from the supplied text.
  • The reported accuracy gains come from the paper's own tests and have not been independently confirmed.
  • The abstract does not state whether V-lens works with models outside the six unified architectures tested.

For security and IT teams

Scope
The findings cover four diffusion-based and two purely autoregressive unified multimodal models, with tuning results reported for BAGEL, SenseNova-U1, and HunyuanImage-3.0.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.