Beyond Accuracy: Auditing Spatial Provenance in Visual Token Pruning for OCR-Critical MLLM Inference
This paper introduces an evidence-risk audit framework for visual token pruning in text-rich MLLMs that reveals how accuracy-based metrics can mask critical failures in spatial provenance, demonstrating that transparent, training-free selectors can achieve comparable accuracy while significantly improving batch speed and memory efficiency without sacrificing the traceability of OCR-critical regions.
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
=== SUMMARY ===
Imagine a world where computers can "see" pictures and answer questions about them, like a super-smart friend who can read a menu from a photo or tell you what's written on a street sign. This is the realm of Multimodal Large Language Models (MLLMs). To do this, the computer breaks an image down into thousands of tiny digital puzzle pieces called "tokens." It then reads these pieces one by one, just like reading words in a book, to figure out what the image is about.
However, looking at a high-resolution photo with lots of text is like trying to read a whole encyclopedia in one sitting; it takes a lot of time and computer power. To make these models faster, scientists have developed a trick called "visual-token pruning." Think of it as a librarian who, before you ask a question, quickly scans a book and throws away the pages they think you won't need, keeping only the most important ones. The goal is to keep the answer correct while making the process much quicker. But here's the catch: just because the librarian thinks they kept the right pages doesn't mean they actually did. If the librarian throws away the page with the specific date you asked about, but the computer guesses the right date anyway because it knows how dates usually look, the answer is right, but the reasoning is broken. This paper investigates exactly that hidden danger.
The Great Token Heist: When the Answer is Right, but the Clues are Gone
The authors of this paper, Feixiang Liu and his team, decided to audit these "librarians" (the pruning methods) with a very specific question: If the computer gives the right answer, did it actually look at the right part of the picture to get it?
They found that in many cases, the answer is "No."
Imagine you are taking a test about a receipt. The question asks, "What is the total price?" The receipt has a tiny, specific number written in a small box. A smart pruning method might throw away 70% of the receipt's image tokens to save time. Surprisingly, the computer might still answer "15.99" correctly. But the authors discovered that in many of these "correct" answers, the computer didn't actually keep the token representing that tiny price box. Instead, it guessed the price based on the shape of the receipt, the font, or just its general knowledge of what receipts usually say.
This is the paper's main discovery: Accuracy alone is a liar. You can have a model that gets the right answer 78% of the time, but if you look closely at where it looked, you'll see it ignored the specific evidence it needed. The authors call this missing link "spatial provenance"—a fancy way of asking, "Can we trace the answer back to the exact spot in the image where the evidence lives?"
The Detective's Toolkit
To prove this, the team built a special "evidence-risk audit." They treated the image like a crime scene.
- The Positive Clue: They picked a specific word or number on the image (like a price) and asked, "Did the computer keep the tokens for this specific spot?"
- The Trap: They also asked questions about things that weren't there (like a fake price) to see if the computer was just guessing or actually checking.
- The Audit: They compared three different ways of picking which tokens to keep:
- Target: The computer tries to guess which tokens are important based on the question.
- Random: The computer picks tokens by rolling a digital dice.
- Grid: The computer picks tokens in a neat checkerboard pattern.
The Shocking Results
The results were eye-opening. When they tested a popular model called Qwen at a 30% retention budget (keeping only 30% of the image tokens):
- The "Target" method kept the answer correct (0.786 accuracy) and managed to keep the specific evidence tokens for about 62% of the questions.
- The "Random" method got the answer right less often (0.739 accuracy) and kept the evidence tokens for only 27% of the questions.
- The "Grid" method was similar to Random, keeping evidence for only 31% of the questions.
Here is the twist: Even though the "Target" method was better, it still missed the evidence in nearly 40% of the cases where it got the answer right! This means the computer was often relying on language patterns instead of the image.
The authors also found that different models play by different rules. What works for one model (like Qwen) doesn't work for another (like LLaVA or InternVL). For example, a method that keeps the evidence safe for LLaVA might actually make it guess more often. There is no "one-size-fits-all" setting.
The Cost of Speed
The paper also looked at the real-world speed benefits. By cutting out 70% of the tokens, they did make the computer faster.
- For the Qwen model, they saw a 4.32 times speedup in processing batches of images.
- They also saved 76.4% of the memory needed to hold the image data.
But the authors warn that this speed comes with a hidden cost. If you are using the model for something critical, like reading a medical prescription or a legal document, guessing the answer because the evidence was thrown away is dangerous. The paper suggests that we shouldn't just report "Accuracy" and "Compression Ratio." We need to report "Spatial Provenance" too—basically, a score that tells us how much of the actual visual evidence survived the pruning.
The Bottom Line
This paper doesn't say that pruning is bad. It says that we need to be smarter about how we measure it. Just because a computer gives the right answer doesn't mean it saw the right thing. The authors propose a new standard: when we compress an image for a computer to read, we must check if the "clues" are still there. If the answer is right but the clues are gone, the system is unreliable, no matter how fast it is.
In the end, the authors suggest that for tasks where reading specific text is crucial (like OCR), we need to keep more of the image than we thought, or at least be honest about how much of the image we actually looked at. They've opened a new door in AI safety, showing that "fast" and "accurate" isn't enough; we also need "traceable."
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.