SkillZip: Evaluation-Free Skill Compression for Self-Evolving Agents by Discovering Reusable Structure
SkillZip is an evaluation-free method for compressing self-evolving agent skills by discovering reusable structural patterns and formalizing them as a typed minimum description-length objective, thereby achieving efficient, maintainable, and generalizable skill representation without the need for costly rollouts or task replay.
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 how to do a complex task, like baking a cake or fixing a computer. At first, you give it a simple list of instructions. But as the robot tries to learn on its own, it starts keeping a diary of every single thing it does. If it burns a cake, it writes down "Don't burn the cake." If it tries a new way to mix the batter and it works, it adds a whole new paragraph about that method. Over time, this diary becomes a massive, messy book. The problem is that the robot has to read the entire book every single time it tries to bake, even if it only needs to remember one tiny rule. This makes the robot slow and confused, like a librarian trying to find a specific book in a library where every shelf is overflowing with duplicate copies of the same story.
This is the world of "self-evolving agents"—computer programs that get smarter by learning from their own mistakes and successes. The big challenge scientists are facing is that these programs are getting too big. They are accumulating so much "redundant" text (repeated rules and copied steps) that they become inefficient. The question is: How do we shrink this giant, messy diary back down into a tiny, perfect instruction manual without deleting any of the important rules? If we just delete the "boring" parts, we might accidentally remove a rare rule that saves the robot from crashing. If we use a test to see what to keep, it takes too long and costs too much money.
Enter SkillZip, a new method that acts like a super-smart editor for these robot diaries. Instead of guessing what to keep or running expensive tests, SkillZip looks at the structure of the instructions. It uses a clever trick: "Explain once, reference many." Imagine you have a recipe that says "Preheat the oven" in three different chapters. SkillZip realizes this is the same rule and moves it to the very top, telling the robot, "Hey, whenever you see a chapter about baking, just look at the top for the oven rule." It finds repeated patterns, groups them into shared "functions," and keeps only the unique differences as special notes.
The researchers found that this approach works incredibly well. They tested it on agents learning to search the web, solve math problems, and use spreadsheets. They discovered that as these agents evolved, their instruction manuals grew to be more than 5 times their original size, mostly because of repeated text. SkillZip was able to shrink these bloated manuals by about 31% on average, making them much faster to read. Even better, it did this without ever needing to run a single test or check the robot's answers against a grading key. In fact, the compressed manuals actually performed just as well, and sometimes even better, than the giant, unedited versions. It turns out that by organizing the knowledge like a well-structured codebook rather than a messy diary, the robot can remember everything it needs to know without carrying the weight of the whole library.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.