Symbal: Detecting Systematic Misalignments in Model-Generated Captions
This paper introduces Symbal, a dual-stage method utilizing foundation models to detect and summarize systematic misalignments between images and MLLM-generated captions, alongside SymbalBench, a large-scale benchmark demonstrating Symbal's superior performance in auditing image-text datasets without requiring access to the underlying models.
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 teaching a robot to describe the world around it. You show it a picture of a dog and ask, "What do you see?" A smart robot might say, "A fluffy dog running in the park." But sometimes, these robot brains get a little too creative or pick up on weird habits. They might look at a picture of a dog on a beach and say, "A dog eating a sandwich," even though there is no sandwich in sight. This happens because the robot learned that "dogs" and "sandwiches" often appear together in its training data, so it assumes they must be there, even when they aren't.
This is the world of Multimodal Large Language Models (MLLMs)—super-smart computers that can see pictures and write sentences about them. They are amazing, but they have a sneaky flaw: they sometimes make the same mistake over and over again. If a robot sees a specific object, like a pacemaker in a medical X-ray, it might always wrongly claim the patient has a heart condition, even if the heart looks fine. These aren't just random typos; they are "systematic misalignments," which is a fancy way of saying the robot has a specific, repeating blind spot. Scientists care about this because if we use these robots to write reports for doctors or describe photos for the blind, we need to know exactly where their brains are glitching so we can fix them.
Enter SYMBAL, a new detective tool created by researchers Maya Varma and her team to catch these repeating robot lies. Think of SYMBAL as a super-organized librarian who doesn't just read one book at a time but scans a library of a million books to find a pattern. Instead of looking at a single picture and its caption, SYMBAL looks at thousands of them at once to answer a big question: "Is there a specific thing in the pictures that makes the robot consistently say something wrong?"
The researchers built SYMBAL to work in two clever steps, like a two-part investigation. First, the tool acts as a text detective. It reads through thousands of robot-generated captions, groups together sentences that say similar things, and asks, "Which of these groups is lying the most?" It uses math to measure how much the text disagrees with the actual image. Once it finds the group of sentences that are most likely to be wrong, it summarizes the lie into a single concept, like "The robot keeps saying 'cardiomegaly' (an enlarged heart)."
In the second step, SYMBAL becomes a visual detective. It takes all the pictures that were paired with that specific lie and asks, "What do these pictures have in common?" It groups the images together and looks for a recurring visual feature. In our example, it might discover that every time the robot said "cardiomegaly," the picture actually contained a "pacemaker" (a medical device). SYMBAL then connects the dots, reporting: "The robot is systematically confusing pacemakers with heart conditions."
To test if their detective was any good, the team created SYMBAL-BENCH, a giant training ground for testing error-detection tools. They didn't just hope for the best; they built 420 different test scenarios using 1.7 million image-text pairs. They took real datasets of natural photos and medical X-rays, then secretly injected specific, fake errors to see if SYMBAL could find them. It was like hiding a specific type of bug in a garden and seeing if the gardener could find it.
The results were impressive. SYMBAL successfully identified these systematic errors in 63.8% of the test datasets. To put that in perspective, the next-best method, which tried to guess the answer in one big leap without breaking it down, only got it right about 17.1% of the time. This suggests that breaking the problem into two smaller, structured steps is a much better strategy than just asking a giant AI to "guess the pattern" all at once.
The team also took SYMBAL out of the lab and into the real world. They used it to audit captions generated by four different popular robot models. In one case, they found that when a robot saw a bus in a picture, it had a strong habit of wrongly claiming there was a handbag on the ground. In another, they discovered that a robot would often invent text on signs in a picture that wasn't actually there. These findings prove that SYMBAL can spot these hidden, repeating errors without needing to know how the robot was built or having access to its internal code.
Ultimately, this paper suggests that SYMBAL is a powerful tool for "auditing" the data we use to train future robots. By finding these systematic misalignments, we can stop robots from learning bad habits before they even start. As the authors note, this is crucial for safety-critical fields like medicine, where a robot confusing a pacemaker for a heart condition could lead to serious mistakes. While SYMBAL isn't a magic wand that fixes everything instantly, it offers a clear, structured way to shine a light on the specific ways our AI companions get it wrong, helping us build better, more reliable models for the future.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.