HiSkill: Empowering LLM Agents with Hierarchical Skill Graphs
This paper introduces HiSkill, a hierarchical skill graph framework that organizes interaction trajectories into a structured graph connecting high-level skills with executable actions, enabling LLM agents to efficiently retrieve task-relevant subgraphs and perform guided execution that outperforms existing methods while reducing token consumption.
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 super-smart robot how to navigate a giant, messy house to find a specific toy. You don't want to give the robot a new, long instruction manual every time it needs to pick up a sock or turn on a light. Instead, you want it to remember "skills" it has learned before, like "how to open a drawer" or "how to carry a cup." This is the world of Large Language Model (LLM) agents—AI systems that can talk and think, but need to learn how to actually do things in the real world (or a digital simulation of it).
For a long time, scientists have tried to help these robots by giving them a list of past experiences, like a diary of things they've done. But there's a problem: these diaries are often just long, flat lists of text. It's like having a library where every book is just a single sentence. If the robot needs to know how to "make a sandwich," it might find a sentence saying "make a sandwich," but it won't know the steps: get bread, get butter, spread butter, put bread together. It also doesn't know what to do if it drops the bread. The robot gets stuck because it has the big idea but lacks the tiny, concrete steps and the backup plans to fix mistakes. This paper asks: How can we organize these memories so the robot can actually use them to solve complex problems without getting confused?
Enter HiSkill, a new method that acts like a master architect for a robot's brain. Instead of a flat list of memories, HiSkill builds a hierarchical skill graph. Think of this graph as a giant, interactive subway map for the robot's mind.
On this map, there are two main types of stations. The big, fancy stations are "Skill Nodes." These are the high-level goals, like "Clean the Kitchen" or "Find the Red Ball." But a station on a subway map isn't useful unless you know which train to take to get there. That's where the smaller stations come in: "AtomicOp Nodes." These are the tiny, concrete actions, like "Grab the sponge," "Turn on the faucet," or "Wipe the counter."
The magic of HiSkill is how it connects these stations. It doesn't just draw a line from "Clean Kitchen" to "Grab sponge." It draws different types of lines (edges) that tell the robot exactly how the pieces fit together:
- Decomposition lines show that "Clean Kitchen" is made of a specific sequence of smaller actions.
- Transition lines show that after you "Grab the sponge," you can usually "Turn on the faucet."
- Recovery lines are like emergency exits; if the robot drops the sponge, these lines point to a "Pick up sponge" action to get back on track.
- Support lines show what extra tools or steps are needed before a skill can start.
When the robot is given a new task, HiSkill doesn't dump its entire brain (the whole graph) into the robot's memory. That would be too much information! Instead, it acts like a smart GPS. It looks at the task, finds the relevant "Skill Station," and then pulls out just the tiny, compact map (a subgraph) needed for that specific journey. It connects the big goal to the tiny steps and the backup plans, all in one neat package.
The paper tested this idea in three different digital worlds: a house where the robot has to move objects (ALFWorld), a website where it has to shop for items (WebShop), and a science lab where it has to run experiments (ScienceWorld). The results were impressive. The HiSkill robot didn't just solve more tasks than the other methods—it solved them much faster and with far less "thinking" required. In fact, it used 78.75% fewer words (tokens) to get the job done compared to the strongest previous method. This suggests that by organizing memories into a structured map with clear connections between big ideas and small actions, robots can become much more efficient and reliable.
The authors also checked what happens if you break the map. If they removed the tiny action stations (AtomicOps), the robot got lost because it only had big ideas. If they removed the special lines (edges) that show how to recover from mistakes, the robot gave up when things went wrong. This proves that the structure itself—the way the big skills and small actions are linked—is the secret sauce.
In short, HiSkill suggests that to make AI agents truly helpful, we shouldn't just give them a pile of notes. We need to give them a well-organized, connected map that shows not just what to do, but how to do it step-by-step, and what to do when things go wrong. It's the difference between handing someone a list of words and handing them a fully illustrated, interactive guidebook.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.