← Latest papers
🤖 AI

Evaluating Visual Prompts with Eye-Tracking Data for MLLM-Based Human Activity Recognition

This paper proposes and evaluates a token-efficient visual prompting strategy that converts high-frequency eye-tracking sensor data into visualization images (such as timelines, heatmaps, and scanpaths) to enable Multimodal Large Language Models (MLLMs) to effectively perform human activity recognition in IoT applications.

Original authors: Jae Young Choi, Seon Gyeom Kim, Hyungjun Yoon, Taeckyung Lee, Donggun Lee, Jaeryung Chung, Jihyung Kil, Ryan Rossi, Sung-Ju Lee, Tak Yeon Lee

Published 2026-04-14
📖 4 min read☕ Coffee break read

Original authors: Jae Young Choi, Seon Gyeom Kim, Hyungjun Yoon, Taeckyung Lee, Donggun Lee, Jaeryung Chung, Jihyung Kil, Ryan Rossi, Sung-Ju Lee, Tak Yeon Lee

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 have a super-smart robot assistant (a Multimodal Large Language Model, or MLLM) that can talk, read, and understand the world. You want to teach this robot to recognize what a person is doing just by looking at their eye movements—like if they are reading a book, playing a game, or browsing the web.

The problem? Eye-tracking data is like a firehose of numbers. It records exactly where a person looks, hundreds of times every second. If you try to feed this raw data to the robot as a giant list of numbers (like "x=10, y=20, x=11, y=21..."), two bad things happen:

  1. It gets overwhelmed: The robot has to read a massive book of numbers just to understand a few seconds of activity. It's like trying to read a dictionary to find the word "cat."
  2. It gets confused: The robot might miss the important patterns because there's too much noise in the middle of the data.

The Solution: "Show, Don't Tell"

The researchers at KAIST and Adobe decided to stop feeding the robot a list of numbers and instead show it pictures. They turned the eye-tracking data into visualizations, similar to how a weather map turns temperature numbers into a colorful image.

They tested three different ways to draw these "eye maps":

  1. The Timeline (The Movie Script): A graph showing where the eyes moved up and down over time. It's like a heartbeat monitor for eyes.
  2. The Heatmap (The Crowd Map): A colorful blob showing where the eyes spent the most time. Bright red spots mean "looked here a lot," while blue spots mean "ignored."
  3. The Scanpath (The Treasure Hunt): A line connecting the dots of where the eyes jumped, showing the exact path the eyes took.

The Experiment: A Taste Test

The researchers fed these pictures to the robot along with a few examples (like showing it a picture of someone reading and saying, "This is reading"). They asked the robot to guess what the person was doing in new pictures.

They tested this on three different groups of people doing various tasks, from simple eye exercises to complex computer work.

What They Discovered

1. Pictures are Cheaper and Faster
Feeding the robot a picture was a huge win.

  • The Cost: Sending a list of numbers for 10 seconds of eye data was like sending a 20-page document. Sending the picture was like sending a single postcard. The robot needed 2 to 20 times fewer "tokens" (the currency of AI processing) to understand the picture.
  • The Result: The robot was often better at guessing the activity when it saw the picture than when it read the numbers.

2. One Size Does Not Fit All
Just like you wouldn't use a map of the whole world to find your way around your house, the researchers found that different pictures work for different activities:

  • Heatmaps were great for activities where where you look matters most (like playing a game where you scan the screen).
  • Scanpaths were better for activities where the order of looking matters (like reading a sentence from left to right).
  • Timelines were good for seeing how fast the eyes were moving.

3. Time Matters
If you show the robot a picture of 10 seconds of eye movement, it might get confused if the activity is complex (like coding). But if you give it a picture of 60 or 100 seconds, the pattern becomes clear. The robot needs a longer "movie" to understand the story of what the person is doing.

The Big Picture

This study is like inventing a new language for robots to talk to sensors. Instead of forcing the robot to do complex math on raw numbers, we give it a visual summary that it can instantly "get."

This is a game-changer for the Internet of Things (IoT). Imagine your smart glasses or smart home system using this method to know if you are stressed, focused, or distracted, without needing to be retrained for every new task. It's a way to make AI smarter, faster, and more efficient by letting it see the data instead of just reading it.

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 →