Source-linked AI summary
Co-Skill: A Collaborative Communication Framework for Skill Evolution
Yilin Ma, Yangi Pan, Weihao Yang, Peixin Zeng, Jiannan Xu, Hao Huang, Wen Xia
TL;DR
Hybrid skill evolution is promising but suffers from blind communication: the cloud lacks visibility into edge execution capability, while the edge does not organize trajectories for cloud analysis. CCF addresses this with capability-aware trajectory and skill structures plus collaborative evolution, reducing token usage while improving success rates across ALFWorld and WebShop.
Problem
Existing hybrid evolution methods face redundant cloud communication and edge execution deficiency because communication between the cloud LLM and edge SLM is blind.
Method
CCF uses a prefix-merged trajectory trie, a progressive skill tree, and collaborative skill evolution to coordinate cloud LLM analysis with edge SLM execution and internalization.
Results
15.6%–41.9% lower SLM+LLM token usage and 25.8%–76.4% higher task success rates are reported across ALFWorld and WebShop versus state-of-the-art hybrid methods.
Takeaways & Limitations
CCF provides an edge-cloud collaborative approach for efficient hybrid skill evolution within the evaluated text-based environments.
Takeaways & Limitations
CO-SKILL is evaluated offline rather than through end-to-end online evolution from live interaction feedback, and extending the approach to vision-language models remains future work.
Abstract
from arXiv · showhide
Agent evolution through skills becomes critical for LLM-based agents to iteratively improve task success rate. Hybrid evolution is a cost-efficient paradigm where a cloud LLM analyzes and generates skills while an edge SLM executes and internalizes them. However, existing hybrid methods, such as SkillRL, still suffer from low success rate and high token usage. We find 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. We thus propose the Collaborative Communication Framework (CCF) to achieve effective edge-cloud evolution. CCF is realized via three techniques: (1) a cloud-aware prefix-merged trajectory trie where the edge compresses trajectories by merging shared prefixes and pinpointing divergence points for efficient cloud analysis, (2) an edge-aware progressive skill tree where the cloud progressively builds a hierarchical skill tree to match edge SLM execution capability, and (3) a collaborative skill evolution scheme upon these two trees that evolves cloud LLM and edge SLM in a separated way to jointly improve task success rate. Experiments across ALFWorld and WebShop show that CCF reduces LLM+SLM tokens by 15.6%--41.9% over state-of-the-art hybrid methods while consistently improving 25.8%--76.4% task success rate.
1 Introduction
Agent skill evolution improves capabilities through accumulated experience, but hybrid approaches remain limited by blind edge-cloud communication. CCF addresses this through mutual-awareness mechanisms and achieves lower token usage with higher task success on ALFWorld and WebShop.
- 1 Introduction: Hybrid evolution combines cloud LLM skill analysis and generation with edge SLM execution and reinforcement-learning internalization, offering a cost-efficient alternative to cloud-only or edge-only evolution.Cloud-only methods are costly, while edge-only methods are limited on complex tasks.
- 1 Introduction: Blind communication is identified as the root cause of hybrid evolution deficiencies: the edge cannot meet cloud analysis needs, while the cloud cannot perceive edge execution capability.The resulting communication mismatch makes uploaded trajectories redundant and cloud guidance difficult for the edge to execute.
- 1 Introduction: A cloud-aware prefix-merged trajectory trie merges shared prefixes and exposes divergence points, reducing redundant trajectory uploads while emphasizing differences for cloud analysis.This directly targets cloud communication redundancy.
- 1 Introduction: An edge-aware progressive skill tree expands from high-level goals to concrete guidance so the cloud can match generated skills to the edge SLM’s execution ability.Collaborative skill evolution then jointly evolves the cloud LLM and edge SLM using the two trees.
- 1 Introduction: 15.6%–41.9% fewer SLM+LLM tokens and 25.8%–76.4% higher task success rates are achieved by CO-SKILL across ALFWorld and WebShop versus state-of-the-art hybrid agents.The reported comparison includes hybrid agents such as SkillRL.
- 1 Introduction: CCF replaces blind communication with mutual awareness, adapting each side’s communication to the other’s capabilities.The framework is designed for bidirectional cloud-edge evolution.
2 Related Work and Motivation
Prior work spans cloud-only, edge-only, and hybrid skill evolution, but hybrid systems can waste tokens and misalign generated skills with edge capabilities. The motivation for CCF is to replace this blind communication with capability-aware coordination.
- 2 Related Work and Motivation: Cloud-only evolution offers strong performance at high cost, whereas edge-only evolution is fast and cheap but generally struggles with complex tasks because of limited reasoning capacity.These trade-offs motivate hybrid evolution as a cost-efficient paradigm.
- 2 Related Work and Motivation: Hybrid evolution combines cloud LLM analysis with edge SLM execution through initialization, trace upload, cloud refinement, and optional edge reinforcement learning.This workflow aims to combine cloud reasoning capacity with cheaper local execution.
- 2 Related Work and Motivation: Figure 1 compares evolution paradigms by their balance of evolution and token consumption, with CCF reported as achieving the best balance through collaborative communication.The caption characterizes the cloud-only/edge-only and state-of-the-art hybrid balances as weak or suboptimal.
- 2 Related Work and Motivation: 15.3%–40.5% of tokens can be wasted by redundantly uploading and reprocessing trajectories with shared prefixes.The cloud must otherwise infer divergence points while reasoning over overlapping content.
- 2 Related Work and Motivation: 33.1%–106.1% more tokens can be required by edge SLMs applying LLM-generated skills on ALFWorld compared with CO-SKILL.Misalignment can instead cause low success rates or inefficient reasoning by the edge SLM.
- 2 Related Work and Motivation: Blind communication leaves the edge unable to provide cloud-friendly trajectories and leaves the cloud unable to generate skills suited to SLM reasoning and execution abilities.The resulting skills can hinder accuracy and reduce edge execution efficiency.
3 Method
CO-SKILL implements Collaborative Communication Framework through bidirectional-aware trees and separated cloud-edge skill evolution. The framework compresses edge uploads around decision forks, progressively matches skills to edge execution capability, and lets each model operate where it excels.
- CO-SKILL coordinates cloud and edge models through a cloud-aware trajectory trie, an edge-aware progressive skill tree, and collaborative skill evolution.These components respectively structure edge-to-cloud communication, cloud-to-edge skills, and joint model or agent evolution.
- 3.1 Overview: Collaborative Communication Framework (CCF): The trie compresses repeated trajectory prefixes and exposes decision forks, allowing the cloud to focus on divergence rather than reprocessing raw trajectories.Uploads are triggered by failure or performance conditions, sufficient recent samples, or a 50,000-token compressed payload, and are asynchronous.
- 3.1 Overview: Collaborative Communication Framework (CCF): The prefix-merged trajectory trie combines shared actions into consensus nodes, branches at diverging actions, and records outcome distributions at each node.Observation delta encoding can further reduce tokens by transmitting only shorter text differences.
- 3.3 Edge-Aware Progressive Skill Tree: The progressive skill tree grows from high-level task guidance toward low-level instructions only where trie evidence shows the edge needs further help.Branches stop growing when accuracy is sufficient, keeping the tree matched to edge capability; the ALFWorld heating example expands through L3 before reaching promising success.
- 3.3 Edge-Aware Progressive Skill Tree: The cloud updates the skill tree by distilling failure paths, contrasting successful and failed paths at forks, and placing corrections at specific nodes.This update process preserves, refines, or adds tree content according to observed edge behavior.
- 3.4 Collaborative Skill Evolution: Collaborative skill evolution separates responsibilities: the cloud refines skill context, while the edge absorbs stabilized skills into its parameters through reinforcement learning.Skill internalization progressively prunes higher-level nodes after convergence, while retaining finer-grained guidance when needed.
4 Evaluation
CO-SKILL is evaluated against cloud-only, edge-only, and hybrid baselines on ALFWorld and WebShop, measuring task success, token efficiency, and component contributions. Across these evaluations, it improves success while reducing overall token usage, with trie, progressive-tree, and collaborative-evolution analyses explaining the gains.
- Overall Accuracy: CO-SKILL reaches 88.9%–90.5% on ALFWorld clean-and-place and cool-and-place, where Cloud-Only achieves only 45% and 35%.On simpler pick-and-place and examine tasks, methods perform competitively, but CO-SKILL converges more quickly.
- Overall Accuracy: On WebShop’s 15-step budget, all methods remain relatively low-performing, yet CO-SKILL approaches the 25.0% cloud baseline and outperforms Skill0 and SkillRL.WebShop evaluates agents on an e-commerce search, browsing, and purchasing task.
- Ablation Studies: The prefix-merged trajectory trie reduces transmission by merging shared prefixes and recording only divergence points across ALFWorld and WebShop episodes.Figure 7 compares merged-trie token counts with naively concatenated raw trajectories.
- Ablation Studies: Progressive skill-tree depth has task-specific optima: further expansion can stop helping or degrade accuracy when detailed instructions overwhelm the SLM.The frozen SLM is evaluated from levels L0 to L5 with cloud LLM and RL disabled.
5 Limitations and Future Work
CO-SKILL is evaluated offline in text-based environments rather than vision-language or fully online settings. Extending it to VLMs and asynchronous online evolution remains future work.
- Offline evaluation uses text-based environments because vision-language benchmarks with rich interaction trajectories are scarce.The authors state that extending CCF-based skill generation, refinement, and execution to VLMs remains future work.
- CO-SKILL does not yet perform end-to-end online evolution from live interaction feedback.A future deployment would require asynchronous RL, where the cloud updates the skill tree while the edge executes tasks; off-policy issues make this non-trivial.
6 Conclusion
CO-SKILL applies collaborative communication to edge-cloud hybrid skill evolution through trajectory compression, progressive skill construction, and joint optimization. On ALFWorld and WebShop, it lowers token usage versus SkillRL while improving task success rate.
- 15.6%–41.9% lower total SLM+LLM token usage and improved task success rate are achieved over SkillRL on ALFWorld and WebShop.The comparison is against the state-of-the-art SkillRL hybrid baseline.
- CCF coordinates edge-cloud evolution through a prefix-merged trajectory trie, progressive skill tree, and collaborative skill evolution mechanism.Together, these components jointly optimize the cloud and edge sides of hybrid skill evolution.