Video reasoning models improve by auditing their own use of visual clues

VIOS, a new method, identifies where a model's predictions depend on video and trains only on those tokens, outperforming standard approaches with a fraction of the compute.

Academic

Rohan Dugad · Mi Luo · Kristen Grauman

University of Texas at Austin

Research Digest·11 November 2026·3 min read

Illustration: Video reasoning models improve by auditing their own use of visual clues

In one line

A model can audit its own visual reliance and focus training on under-utilized visual tokens, improving video reasoning with far fewer resources.

Before you read this

Three ideas this paper builds on

on-policy distillationA training method where a student model learns from a teacher using its own current policy's outputs, often improving reasoning and instruction following.128 papers →evidence blindnessHaving access to relevant evidence does not prevent an LLM agent from making incorrect or unsafe decisions.30 papers →privileged informationInformation available only during training that improves performance but is not accessible at inference.80 papers →

Read it as

Video reasoning often depends on just a few key moments in a video. Current training methods treat every part equally, wasting time and resources.

A technique called Visually-Aware On-Policy Self-Distillation (VIOS) lets a model check how much it relies on the video at each step. It then focuses training on the steps where the model ignores important visual cues.

VIOS improves performance across many benchmarks while requiring 9x fewer training steps and 111x fewer generated tokens than other methods. This makes video reasoning much more efficient.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

on-policy distillation evidence blindness privileged information contrastive alignment chain of thought audit workflow