Meta Context Engineering via Agentic Skill Evolution
arxiv:2601.21557
Agents, open-endedness, AGI · 2026-01-29
Complete cleaned Markdown · Source Markdown including bibliography
Read in parts
Contents
- 1 Introduction
- 2 Background and Related Work
- 2.1 Agentic Context Engineering
- 2.2 Evolutionary Computation with LLMs
- 3 Meta Context Engineering
- 3.1 Problem Formulation
- 3.2 Meta-Level: Agentic Skill Evolution
- 3.3 Base-Level: Fully Agentic Context Optimization
- 3.4 Algorithmic Orchestration
- 4 Experiments
- 4.1 Experimental Setup
- Tasks and Datasets.
- Baselines.
- Models.
- MCE Instantiation.
- Evaluation Settings.
- 4.2 Main Results
- 4.2.1 Context Performance
- 4.2.2 Context Adaptability, Efficiency, and Transferability
- 4.2.3 Training Efficiency
- 4.3 Ablation Studies
- Ablating the Bi-level Design ( Table 3 ).
- Ruling Out Model Confounds ( Table 4 ).
- 5 Conclusion and Discussion
- Appendix A Experimental Details
- A.1 FiNER
- A.2 USPTO-50k
- A.3 Symptom2Disease
- A.4 LawBench
- A.5 Aegis2
- Appendix B Prompts
- Appendix C Learned Skills
- Appendix D Case Studies and Analysis
- D.1 FiNER
- D.1.1 Task Characteristics and Baseline Limitations
- D.1.2 Skill Evolution
- D.1.3 Context Evolution
- D.2 USPTO50k
- D.2.1 Task Characteristics and Baseline Limitations
- D.2.2 Skill Evolution
- D.2.3 Context Evolution
- D.3 Symptom2Disease
- D.3.1 Task Characteristics and Baseline Limitations
- D.3.2 Skill Evolution
- D.3.3 Context Evolution
- D.4 LawBench
- D.4.1 Task Characteristics and Baseline Limitations
- D.4.2 Skill Evolution
- D.4.3 Context Evolution
- D.5 Aegis2.0
- D.5.1 Task Characteristics and Baseline Limitations
- D.5.2 Skill Evolution
- D.5.3 Context Evolution
- Appendix E Utility Functions
- Instructions for reporting errors