Video RLVR induces shortcut amplification via perceptual uncertainty collapse

The authors demonstrate that outcome-level optimization in video reinforcement learning progressively reduces reliance on visual evidence, and propose Bifocal Self-distillation Policy Optimization (BSPO) to mitigate this.

Independent

Jiafeng Liang · Xianhao Ou · Tao Ren · Rongchuan Mu · Runxuan Liu · Dandan Tu · +3 more

Research Digest·28 October 2026·2 min read

Illustration: Video RLVR induces shortcut amplification via perceptual uncertainty collapse

In one line

Video RLVR training causes shortcut amplification by collapsing perceptual uncertainty, favoring text priors over visual evidence.

Before you read this

Two ideas this paper builds on

reinforcement learning with verifiable rewardsReinforcement learning that uses rewards from objectively checkable signals like correct answers or verification feedback.131 papers →counterfactual reasoningComparing actual outcomes with imagined alternatives to estimate the effects of different actions or conditions.82 papers →

Read it as

When training AI models on video tasks using rewards for correct answers, models often learn text shortcuts instead of understanding the video. They guess based on common text patterns even when the video contradicts them.

Researchers found this comes from a collapse of uncertainty that ignores visual evidence. They created Bifocal Self-distillation Policy Optimization (BSPO), which uses two signals: comparing good and bad answers, and penalizing predictions that don't change when visual clues are weakened.

BSPO improves video reasoning by making models pay attention to both reasoning and visual details. This results in more reliable AI that truly understands videos instead of taking shortcuts.

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

reinforcement learning with verifiable rewards perceptual uncertainty collapse shortcut amplification knowledge distillation counterfactual reasoning