From Multi-Agent to Single-Agent: When Is Skill Distillation Beneficial?
Binyan Xu, Dong Fang, Haitao Li, Kehuan Zhang
cs.AI
Submitted: 2026-08-19
Updated: 2026-08-20
Comments: 35 pages, 15 figures, 17 tables
Code: https://github.com/Tencent/AdaSkill
Project page: https://adaptive-skill.github.io/Codehttps://github.com/Tencent/AdaSkill
License: http://creativecommons.org/licenses/by/4.0/
The gist: Multi-agent systems (MAS) for structured data-science tasks externalize analytical control through workflows spanning stages, tools, shared state, verification, and repair.
Terminology
Abstract
Multi-agent systems (MAS) for structured data-science tasks externalize analytical control through workflows spanning stages, tools, shared state, verification, and repair. Distilling such workflows into a single-agent skill can reduce orchestration overhead, but it remains unclear which workflow components should cross the control boundary. We distinguish capability resources, which expand what an agent can do, from pipeline guidance, which constrains which solutions it explores. On the same causal-estimation instances, adding task-qualified source pipeline guidance to a capability-matched skill changes normalized utility by +19.6 points under method-selection accuracy but-10.3 points under numerical error. To explain this reversal, we introduce Behavior-Outcome Freedom (F), a pre-synthesis diagnostic of signed behavior-outcome rank mismatch, and formalize its candidate-conditional role through Signed Anchor-Rank Transfer. Motivated by this mechanism, we propose AdaSkill, which preserves validated capability resources, removes runtime orchestration, and conditionally inherits pipeline guidance using a calibrated rule over F. Across 16 capability-matched interventions, the native-scale Full-minus-Discard effect decreases across the continuous F scale (r = -0.80, p < 0.001), while a 15-treatment atomic sweep localizes the reversal to pipeline guidance. Across 11 datasets spanning four structured data-science task families, AdaSkill combines strong task performance with substantially lower deployment overhead.
Sources
- GPT-4 Technical Report
- SMAGDi: Socratic Multi Agent Interaction Graph Distillation for Efficient High Accuracy Reasoning
- EvoSkill: Automated Skill Discovery for Multi-Agent Systems
- APEX-SQL: Talking to the data via Agentic Exploration for Text-to-SQL
- DS-Agent: Automated Data Science by Empowering Large Language Models with Case-Based Reasoning
- Distilling the Knowledge in a Neural Network
- When Single-Agent with Skills Replace Multi-Agent Systems and When They Fail
- AgentArk: Distilling Multi-Agent Intelligence into a Single LLM Agent
- Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills
- FELA: A Multi-Agent Evolutionary System for Feature Engineering of Industrial Event Log Data
- Scaling Large Language Model-based Multi-Agent Collaboration
- Voyager: An Open-Ended Embodied Agent with Large Language Models
- Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents
- SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
- When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting
- Contextual Agentic Memory is a Memo, Not True Memory
- LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception
- From Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Data-Free Online Backdoor Defense
- Rethinking the Value of Multi-Agent Workflow: A Strong Single Agent Baseline
- AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Agents' Last Exam
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems