Research map / Large language models / Agents, code and math
Meta Context Engineering via Agentic Skill Evolution
Haoran Ye, Xuning He, Vincent Arak, Haonan Dong, Guojie Song
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| evolve skills across iterations instead of skill-less base agent (ablation) | MCE · FiNER (offline) | accuracy (%): 73.0 -> 75.0 | Improved |
| evolve skills versus a single fixed skill from the task spec (ablation) | MCE · FiNER (offline) | accuracy (%): 71.0 -> 75.0 | Improved |
| add a fixed skill to the base agent (online) (ablation) | MCE · FiNER (online) | accuracy (%): 67.0 -> 68.0 | No clear effect |
| add a fixed skill to the base agent (offline) (ablation) | MCE · FiNER (offline) | accuracy (%): 73.0 -> 71.0 | Hurt |
| fully agentic base agent without any manual scaffolding versus ACE (ablation) | MCE (w/o skills) · FiNER | accuracy (%): 71.0 -> 73.0 offline; 64.0 -> 67.0 online | Improved |
| use MiniMax M2.1 as ACE reflector/curator instead of DeepSeek V3.1 (ablation) | ACE · FiNER | final accuracy (%): 70 -> 67 | Hurt |
| replace ACE's additive curation with MCE's agentic batch-level optimization | context learning · FiNER | training time (hours): 25.8h -> 1.9h (13.6x faster) | Improved |
| use MCE instead of ACE for rollout efficiency | context learning · FiNER | rollouts to reach training accuracy: 450 rollouts for 95% vs ACE peak 94% at 2169 | Improved |
| use MCE contexts instead of ACE contexts for transfer to Llama3.3-70B | transferred context · LawBench (transfer to Llama3.3-70B) | micro-F1: ACE 0.19 (base 0.24) vs MCE 0.27 | Improved |
| use MCE instead of ACE for the Symptom2Disease online setting | context learning · Symptom2Disease (online) | accuracy (%): 62.3 -> 76.4 | Improved |
| ACE reflection-curation on a task where raw examples suffice | ACE · Symptom2Disease (offline) | accuracy (%): ICL 84.4 vs ACE 79.2 | Hurt |
| MCE versus GEPA on a lightweight generator | MCE · Aegis2.0 (Qwen3-8B) | F1: 0.76 -> 0.80 | Improved |
Closest papers by meaning
- SkillOpt: Executive Strategy for Self-Evolving Agent Skills (NeurIPS 2026 · similarity 0.88)
- MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.87)
- Kimi-Dev: Agentless Training as Skill Prior for SWE-agents (ICLR 2026 · similarity 0.87)
- From History to State: Constant-Context Skill Learning for LLM Agents (NeurIPS 2026 · similarity 0.87)
- Meta$^n$: Recursive Self-Improvement through Emergent Depth (NeurIPS 2026 · similarity 0.87)
- MCP-AgentBench: Evaluating Real-World Language Agent Performance with MCP-Mediated Tools (AAAI 2026 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
