Reduction in total SLM+LLM token usage compared to state-of-the-art hybrid methods.
Consistent improvement in task success rates across ALFWorld and WebShop benchmarks.
Wasted tokens saved by eliminating cloud communication redundancy.
Reduction in per-episode token cost relative to SkillRL after training stabilizes.
Background
Agent evolution through skills is critical for LLM-based agents to iteratively improve task success rates. Hybrid evolution (cloud LLM generates skills, edge SLM executes) is a promising cost-efficient paradigm.
The Problem
Existing hybrid methods (e.g., SkillRL) suffer from low success rates and high token usage. This stems from "blind communication": the cloud cannot perceive the edge's execution capability, while the edge does not understand the cloud's analysis needs.
Our Solution
We propose the Collaborative Communication Framework (CCF) to achieve effective edge-cloud evolution. Experiments across ALFWorld and WebShop show that CCF reduces LLM+SLM token usage by 15.6%-41.9% while consistently improving task success rates by 25.8%-76.4%.
Co-Skill includes three major components: a cloud-aware prefix-merged trajectory trie that compresses edge-to-cloud traffic; an edge-aware progressive skill tree that organizes cloud-to-edge skills; and a collaborative skill evolution method that jointly evolves edge-cloud models.
The Collaborative Communication Framework (CCF) is realized via three key techniques:
Compresses trajectories by merging shared prefixes and pinpointing divergence points.
Progressively builds a hierarchical skill tree to match edge SLM execution capability.
Evolves cloud LLM and edge SLM in a separated way to jointly improve task success rate.
The edge agent compresses trajectories by merging shared prefixes and pinpointing divergence points. This eliminates redundant uploads and surfaces key decision forks for efficient cloud analysis.
The cloud progressively builds a hierarchical skill tree to match the edge SLM's execution capability. It expands the tree only when the edge genuinely needs finer sub-steps.
Operating atop the two trees, the cloud LLM and edge SLM evolve separately. The cloud maintains a hierarchical skill library for persistence, while the edge learns improved skills with an optional RL module for internalization.
We evaluate Co-Skill across standard interactive agent benchmarks: ALFWorld (household environment) and WebShop (e-commerce).
Across ALFWorld and WebShop workloads
Superior Task Success Rate: Co-Skill consistently surpasses edge baselines (e.g., Skill0, SkillRL) and approaches—or even exceeds—cloud-only evolution performance on complex multi-step tasks.
Average SLM+LLM token consumption
High Token Efficiency: Despite requiring comprehensive cloud analysis, total SLM+LLM token usage is reduced by up to 41.9%. This is achieved by the prefix-merged trie eliminating redundant transmissions and the cloud retrieving stored skills for recurring tasks.
@article{ma2025coskill,
author = {Ma, Yilin and Pan, Yangi and Yang, Weihao and Zeng, Peixin and Xu, Jiannan and Huang, Hao and Xia, Wen},
title = {Co-Skill: A Collaborative Communication Framework for Skill Evolution},
journal = {arXiv},
year = {2026},
}