Learned Curation Halves the Skill Bank, Keeps the 9.69-Point Gain

Notes on CODESKILL: Learning Self-Evolving Skills for Coding Agents (arXiv:2605.25430) — Yanzhou Li, Yiran Zhang, Xiaoyu Zhang, Xiaoxia Liu, Yang Liu · May 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

CODESKILL treats skill-bank curation as something a model learns, not a heuristic bolted onto a frozen agent. A management policy — trained through a three-stage curriculum with GRPO — watches coding-agent trajectories and decides whether to extract a new procedural skill (a high-level task strategy, or a local event-driven reaction to something like a recurring command failure), revise an existing one against new evidence, or merge and drop entries during maintenance. Three signals shape the reward: a rubric-based reward from an LLM judge scoring grounding, reusability, and actionability; sparse execution feedback on whether injecting the candidate skill actually raised the frozen downstream agent's pass rate on held-out tasks; and an alignment check on whether the agent's behavior matched what the skill claimed to trigger. Training uses a group size of six generations per prompt, a quality-reward weight of 0.25, and a KL coefficient of 0.02, optimized against a frozen Qwen3.5-35B-A3B coding agent.

Across EnvBench (Python and Java), SWE-Bench Verified, and Terminal-Bench 2, CODESKILL lifts average pass rate 9.69 points over a no-memory baseline and 4.01 points over the strongest prompt-based or memory baseline tested per benchmark (a fixed-prompt "Prompt Skill Mgmt." approach on EnvBench and SWE-Bench Verified, "Subtask Memory" on Terminal-Bench 2). EnvBench-Python shows the widest gap — pass rate nearly triples over no-skill, 18.60 versus 6.98. Maintenance is doing real work, not just bookkeeping: extraction and evolution alone would grow the bank to 1,252 skills; running the full lifecycle with merge/drop maintenance compacts that to 676 — roughly half — without giving back the gain. Swap the frozen downstream agent for GPT-5.4-mini and the improvement over no-skill still holds at 8.93 points, evidence the learned skills capture something more portable than one model's prompting quirks.

Key numbers

Average pass-rate gain vs. no-skill baseline	+9.69 pts
Gain vs. strongest prior baseline	+4.01 pts
EnvBench-Python pass rate (CODESKILL vs. no-skill)	18.60 vs. 6.98
Skill bank size after maintenance (from 1,252 unpruned)	676 skills
Gain when swapped to a different frozen agent (GPT-5.4-mini)	+8.93 pts

What this means for skill authors

The result worth generalizing past this exact RL recipe: skill-bank maintenance works better as a judged decision than as a fixed rule — aging entries out after N days, deduping by cosine similarity. That's what keeps a collection useful instead of either bloating or purging good entries alongside stale ones. CODESKILL's maintenance step doesn't just accumulate: extraction and evolution alone would double the eventual bank size before pruning kicks in, and the system is actively deciding whether a new candidate duplicates something already present or captures a genuinely new failure mode.

For teams building or consuming a shared skill library, the load-bearing detail is that quality gets scored on grounding, reusability, and actionability as separate axes, then checked against whether the skill was actually followed and whether following it helped. That's close in spirit to why a catalog benefits from vetting a skill on more than one axis before shipping it to users — a skill that merely runs without errors and a skill that's specific enough to trigger correctly and general enough to transfer are different bars, and conflating them is how bloat starts.

The cross-model transfer number (+8.93 on GPT-5.4-mini versus +9.69 on the training-time agent) matters most for anyone shipping skills across model versions or vendors: skills that only help the exact model that authored them stop being an asset the moment that model gets upgraded or swapped. A learned skill bank that keeps most of its lift after an agent swap is evidence the skills encode task strategy rather than model-specific prompting habits — the property that actually matters for a skill meant to be installed by someone other than its author.

Skills related to this research

fine-tuning-with-trl This skill teaches post-training techniques for aligning language models to human preferences. It covers supervised fine-tuning, direct preference optimization (DPO), and online reinforcement learning methods like RLOO and GRPO, with complete workflows and practical examples.★ 221,503 fine-tuning-with-trl Fine-tuning with TRL provides post-training methods to align language models with human preferences through multiple approaches. Train models on instruction data with SFT, optimize for preference alignment via DPO without a separate reward model, or run full RLHF pipelines combining supervised fine-tuning, reward modeling, and PPO optimization. GRPO offers memory-efficient online reinforcement learning for resource-constrained setups.★ 11,165 Unsloth Training Skill This skill teaches efficient LLM fine-tuning via Unsloth, covering GRPO (reinforcement learning with reward functions) and SFT (supervised training). Learn FP8 training for 60% VRAM reduction, automatic packing for 2–5x speedup, vision model tuning, mobile deployment via QAT, Docker containerization, and export to GGUF and other formats.★ 26 Unsloth Training Unsloth Training accelerates LLM fine-tuning through GRPO (reinforcement learning with reward functions) and SFT (supervised learning with input-output pairs). It cuts VRAM usage by 60% with FP8 training, speeds up mixed-length datasets via automatic packing, and supports vision model tuning, mobile deployment via ExecuTorch, and export to GGUF, Ollama, and vLLM.★ 8

Related notes

References

Li, Zhang, Zhang, Liu, Liu — CODESKILL: Learning Self-Evolving Skills for Coding Agents, arXiv:2605.25430 (2026)