← Latest papers
💻 computer science

Dynamic Token Compression for Efficient Video Understanding through Reinforcement Learning

The paper proposes SCORE, a reinforcement learning-based framework that utilizes surprise-augmented state representations and a curriculum learning strategy to achieve adaptive token compression, significantly accelerating video understanding while maintaining near-perfect performance.

Original authors: Shida Wang, YongXiang Hua, Zhou Tao, Haoyu Cao, Linli Xu

Published 2026-03-30
📖 4 min read☕ Coffee break read

Original authors: Shida Wang, YongXiang Hua, Zhou Tao, Haoyu Cao, Linli Xu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to explain a two-hour movie to a friend who has a very short attention span and can only listen to a few sentences before getting overwhelmed.

If you try to describe every single frame of the movie—every leaf on a tree, every cloud in the sky, every second of a character standing still—your friend's brain will crash. They will get confused, forget the beginning of the story, and miss the important plot twists. In the world of Artificial Intelligence, this is called "Context Rot." The AI is drowning in too much information, and the "noise" (redundant, boring frames) drowns out the "signal" (the exciting action).

This paper introduces a new system called SCORE to solve this problem. Think of SCORE as a super-smart, fast-forward button that doesn't just skip randomly, but knows exactly what to keep and what to throw away.

Here is how SCORE works, broken down into simple concepts:

1. The Problem: The "Information Flood"

Modern AI models (like the ones that watch videos) are like sponges. When you feed them a video, they turn every frame into thousands of tiny data points called "tokens."

  • The Issue: A 10-minute video might generate 50,000 tokens. But 80% of them are boring! They are just the sky, a wall, or a person standing still.
  • The Result: The AI gets "clogged up." It spends all its energy processing the boring stuff and forgets the important stuff. It's like trying to find a needle in a haystack, but the haystack is made of 99% other needles.

2. The Solution: The "Surprise Detective"

SCORE is a smart filter that sits between the video camera and the AI brain. Its job is to decide which tokens to keep and which to delete before the AI even sees them.

To do this, SCORE uses a trick called "Surprise-Augmented State."

  • The Analogy: Imagine you are watching a movie. If the screen shows a static image of a wall for 10 seconds, your brain gets bored and stops paying attention. But if a cat suddenly jumps onto the table, your brain goes, "Whoa! Surprise!" and focuses immediately.
  • How SCORE does it: It compares every frame to the one before it.
    • Frame A: A bird sitting on a branch.
    • Frame B: The bird is still sitting there. (Difference = 0. Boring. Delete.)
    • Frame C: The bird flies away. (Difference = Huge. Surprise! Keep.)
  • By looking for these "surprises" (changes), SCORE knows exactly when something important is happening and keeps those tokens, while deleting the boring, repetitive ones.

3. The Training: Learning by Playing a Game

How does the AI learn to be this good at filtering? It uses Reinforcement Learning, which is like training a dog with treats.

  • The Game: The AI tries to compress the video.
  • The Reward: If it deletes too much and the AI brain can't answer the question correctly, it gets a "bad grade" (penalty). If it deletes the boring stuff but keeps the important stuff, and the AI brain gets a perfect score, it gets a "treat" (reward).
  • The Curriculum: At first, the AI trains on "fake videos" made of static pictures (easy to learn the basics of what changes). Once it masters that, it moves on to real, messy videos with real motion. This step-by-step learning ensures it doesn't get confused by the chaos of the real world.

4. The Result: Speed and Smarts

The paper shows that SCORE is a game-changer:

  • Speed: It makes the AI 16 times faster at processing videos. It's like going from a dial-up internet connection to fiber optics.
  • Smarts: Even when it throws away 90% of the video data (keeping only 10%), the AI still understands the video almost perfectly. In fact, by removing the "noise," the AI sometimes understands the video better than if it had seen the whole thing!

Summary

SCORE is like a personal editor for AI. Instead of forcing the AI to read the entire encyclopedia of a video, SCORE highlights the most exciting chapters, skips the boring filler, and hands the AI a concise, high-quality summary. This allows the AI to think faster, remember better, and understand long videos without getting overwhelmed.

Drowning in papers in your field?

Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.

Try Digest →