AI 代理與開發工具
SKILLER Treats SKILL.md as an Optimizable Policy, Helping Qwen3.5-9B Reach 82.8% on Software Engineering Tests
Rather than updating the small model’s weights, SKILLER has a stronger model repeatedly revise text-based skills using execution traces and verifier feedback. Qwen3.5-9B’s three-run average on SWE-Skills-Bench rose from 26.2% without skills to 82.8%, although skill generation still relies on GPT-5.4 and successful reference trajectories.
The research team has open-sourced SKILLER, which treats agent skills as natural-language policies that can be searched through reinforcement learning, rather than as fixed prompts or model parameters. The target Qwen3.5-4B/9B model executes the current `SKILL.md` in a tool environment; an official verifier returns the success rate, test results, and error messages, after which GPT-5.4 serves as both critic and actor. The critic compares failed trajectories, successful reference trajectories, and historical memory to identify the earliest causal divergence. The actor then makes localized revisions to the skill using only four operations—Insert, Replace, Create, and Delete—and can also create deterministic helper programs when necessary. No backpropagation is performed through the execution model during this process.
Under a five-step optimization schedule, Qwen3.5-9B scored 82.8% across 117 SWE-Skills-Bench cases, outperforming the 26.2% baseline without skills, SkillX at 58.8%, and Manus skills at 62.4%. The 4B version improved from 17.6% to 66.7%. Performance on GAIA’s held-out samples also rose from 37.0% to 49.59%, suggesting that the method does more than memorize individual training cases. The key technical idea is to shift expensive computation to the offline skill-construction stage, while deployment can still use smaller local models and reviewable text artifacts.
However, this is not a way to avoid frontier models altogether: the generation stage still requires GPT-5.4, verifiers, and training-specific reference trajectories. The paper’s cost analysis also excludes tokens consumed by subsequent evaluations of the small models. Every reported score is only an average of three runs, and SkillsBench currently includes just 26 single-skill tasks. The public repository provides the controller, prompts, benchmark adapters, and skill snapshots, but excludes data restricted by upstream terms, complete trajectories, and model weights. The SKILLER code is also released under the PolyForm Noncommercial License. Future work should test whether these skills remain effective across different models, runtime versions, and previously unseen tool interfaces.