skillfed

Evolve the context-engineering skill itself, not just the context — up to 54% better than fixed pipelines like ACE and GEPA

Notes on Meta Context Engineering via Agentic Skill Evolution (arXiv:2601.21557) — Haoran Ye, X. He, Vincent Arak, Haonan Dong, Guojie Song · January 2026

Note published · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

MCE reframes context engineering as a two-level search rather than a fixed pipeline. Instead of hand-designing the harness that decides how context gets built — the generation-reflection-curation loops and itemized-list schemas behind methods like ACE, or the brevity-biased prompt rewrites behind GEPA — a meta-level agent evolves the harness itself: executable "skills" that combine natural-language methodology, scripts, and validation logic. It refines these through agentic crossover, a deliberative reasoning pass over the full history of prior skills, their executions, and validation scores, rather than a fixed recombination rule. A separate base-level agent then executes whatever skill the meta-agent hands it, reading training rollouts and writing the actual context as files and code instead of filling in a predefined schema. The two levels co-evolve under a simple (1+1) evolution strategy, tested across five unrelated domains: XBRL entity tagging (FiNER), retrosynthesis prediction (USPTO-50k), symptom-to-disease classification, Chinese criminal-charge prediction (LawBench), and prompt-safety classification (AEGIS2). DeepSeek-V3.1 is the model under test (Qwen3-8B for the safety task), MiniMax M2.1 does the evolving, and the field it's measured against includes ICL, MIPROv2, GEPA, Dynamic Cheatsheet, and Agentic Context Engineering (ACE).

MCE sweeps all five benchmarks in the offline setting, posting an 89.1% average relative gain over the frozen base model against 70.7% for ACE, the next best. Online, where context has to be built from a single pass over the test stream, the gap holds at 74.1% versus ACE's 41.1%. Pooled across every baseline and setting, MCE's edge ranges 5.6-53.8% in relative improvement, averaging 16.9%. It also escapes the length habits baked into fixed pipelines: GEPA stays stuck near 1-2K tokens, ACE bloats to roughly 80K after five epochs, while MCE ranges from 1.5K to 86K tokens depending on what a task actually needs. On FiNER, it trains 13.6x faster than ACE (1.9 hours versus 25.8) and reaches higher training accuracy (95% versus ACE's peak of 94%) on 4.8x fewer rollouts (450 versus 2,169). Contexts trained on DeepSeek-V3.1 also degrade less when handed down to smaller models — 4-7% less relative drop than ACE's contexts — and on two benchmarks the resulting general-model-plus-context beats a dedicated specialist outright: 0.70 vs. 0.56 F1 on LawBench, 0.80 vs. 0.72 F1 against Llama Guard 3 8B on AEGIS2.

Key numbers

Offline avg. relative gain over base model89.1%
Improvement over best CE baselines (mean, range)16.9% (5.6-53.8%)
Training speedup vs. ACE (FiNER)13.6×
Rollout efficiency vs. ACE (FiNER)4.8× (450 vs. 2,169 rollouts)
Strong-to-weak transfer degradation vs. ACE4-7% less relative drop

Skills related to this research

Clerk Android Clerk Android guides you through integrating Clerk authentication into native Android projects using the clerk-android SDK. Choose between prebuilt AuthView and UserButton components for rapid setup, or build custom API-driven flows for complete control over your auth experience.★ 63 Clerk Chrome Extension Patterns This skill covers authentication patterns for Chrome extensions built with Clerk, including popup and side panel UIs, OAuth delegation via syncHost, and service worker token management. Learn which auth methods work in each extension context, configure stable extension IDs, and handle session syncing with your web app.★ 63
Evolutionary Context Search for Automated Skill Acquisition Large Language Models cannot reliably acquire new knowledge post-deployment -- even when relevant text resources exist, models fail to transform them into… Sun et al. · similarity 0.87 Skill-MAS: Evolving Meta-Skill for Automatic Multi-Agent Systems Large Language Model (LLM)-based automatic Multi-Agent Systems (MAS) generation has become a crucial frontier for tackling complex tasks. However, existing… Lin et al. · similarity 0.85 Agent-World: Scaling Real-World Environment Synthesis for Evolving General Agent Intelligence Large language models are increasingly expected to serve as general-purpose agents that interact with external, stateful tool environments. The Model Context… Dong et al. · similarity 0.84

Related notes

References

  1. Ye, H., He, X., Arak, V., Dong, H., & Song, G. (2026). Meta Context Engineering via Agentic Skill Evolution. arXiv:2601.21557.
  2. Zhang, Q., Hu, C., Upasani, S., Ma, B., Hong, F., Kamanuru, V., Rainton, J., Wu, C., Ji, M., Li, H., et al. (2026). Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models. The Fourteenth International Conference on Learning Representations (ICLR).
  3. Agrawal, L. A., Tan, S., Soylu, D., Ziems, N., Khare, R., Opsahl-Ong, K., Singhvi, A., Shandilya, H., Ryan, M. J., Jiang, M., et al. (2025). GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning. arXiv:2507.19457.
  4. Suzgun, M., Yuksekgonul, M., Bianchi, F., Jurafsky, D., & Zou, J. (2025). Dynamic Cheatsheet: Test-Time Learning with Adaptive Memory. arXiv:2504.07952.
  5. Opsahl-Ong, K., Ryan, M. J., Purtell, J., Broman, D., Potts, C., Zaharia, M., & Khattab, O. (2024). Optimizing Instructions and Demonstrations for Multi-Stage Language Model Programs. EMNLP.