Demystifying Entropy-based Selection for Chain-of-Thought Compression in Large Reasoning Models
This paper challenges the efficacy of entropy-based pruning for compressing Chain-of-Thought reasoning, demonstrating that such heuristics offer no advantage over random selection and that task-critical information is distributed across the entire reasoning chain rather than concentrated in specific tokens identifiable by entropy metrics.
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 teach a brilliant but chatty robot how to solve a tricky puzzle. To get the right answer, the robot doesn't just blurt out a solution; it talks to itself first, writing down a long list of thoughts, calculations, and "aha!" moments. This internal monologue is called "Chain-of-Thought." It's like a detective writing down every clue, even the ones that seem obvious, before catching the criminal. While this method makes the robot much smarter, it also makes it very slow and hungry for computer memory, because the list of thoughts can get incredibly long. Scientists are now trying to figure out how to shorten this list without losing the detective's genius. They want to know: can we just delete the boring parts of the story and keep only the exciting, important bits?
For a while, many researchers believed they had found a magic filter. They thought that by looking at how "surprised" the robot was by its own next word (a concept called "entropy"), they could tell which thoughts were crucial and which were just filler. The idea was that if the robot was very confident about what it was going to say next (low surprise), that part was probably just boring repetition and could be cut. If the robot was unsure (high surprise), that was the moment of real thinking and should be kept. It sounded like a perfect way to shrink the robot's diary down to a few pages while keeping all the good stuff. But is this magic filter actually real, or is it just a lucky guess?
This paper is a reality check for that idea. The authors, a team of curious scientists, decided to test this "entropy filter" against a much simpler, dumber method: just deleting thoughts at random. They treated the robot's long reasoning chains like a messy room and tried to tidy it up using two different strategies. The first strategy was the "smart" one, using the entropy filter to pick what to keep. The second was the "random" one, where they just grabbed a handful of sentences or words to keep, ignoring the content entirely. They tested this on a variety of robots (different AI models) and gave them different types of homework, including math problems, logic puzzles, and science questions.
The results were a bit of a plot twist. When the scientists looked at whole sentences, the "smart" entropy filter was no better than the random guess. In fact, the random method often did just as well, or even better, at keeping the robot's performance high. It turns out that the "surprise level" of a sentence isn't a reliable map to where the important information is hiding. The authors suggest that the reason some previous studies thought entropy was working might be because they were mostly testing on math problems, where the "important" words are often just numbers.
When they zoomed in to look at individual words (tokens) instead of whole sentences, a strange pattern emerged. On math tests, keeping the words with low surprise did seem to help. But the authors dug deeper and realized this wasn't because those words were "thinking" words. It was because, in math problems, the low-surprise words were almost always numbers (like "2", "7", or "5"). Since numbers are predictable in math, they have low entropy. The filter wasn't finding the "smart" parts; it was just accidentally picking the numbers. When the scientists removed the numbers from the mix, the low-entropy filter stopped working and became just as bad as the random guess.
To prove this, they used a special trick called "activation patching." Imagine you take the robot's brain activity from the full, long version of the story and paste it into the short, compressed version. This lets them see if the missing context was actually the problem. They found that even with this brain-boosting trick, the entropy filter still couldn't beat the random method unless it was specifically keeping the numbers in math problems. On non-math tasks, like logic puzzles, the entropy filter failed completely, performing no better than chance.
So, what's the takeaway? The paper suggests that the idea of using "surprise" (entropy) to automatically find and keep the most important parts of a robot's thinking process is mostly a myth. The semantic content—the actual meaning and logic of the reasoning—isn't concentrated in a few special, low-surprise words that a simple rule can find. Instead, the important information is spread out across the whole chain of thought. While keeping numbers helps with math, there is no universal "magic filter" that can compress a robot's reasoning chain without losing its smarts. The best way to shrink the list might just be to keep more of it, or to find a smarter way to cut it that doesn't rely on guessing which words are boring.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.