SKILLER Generates Agent Skills Tailored to Small Local Models, Propagating All RL Signal in Natural Language — 4.3 to 20.4 Point Gains on Qwen3.5-9B/4B

arXiv (via HuggingFace Daily Papers)·medium signalXBlueskyLinkedInCopy link

arXiv 2608.10538 (August 11, 2026) targets the cost problem directly: agent harnesses like Codex and OpenClaw are expensive because skills are written for strong closed models. SKILLER uses a strong model as both actor and critic, treats the small-model agent system as the environment, and propagates every reinforcement signal as natural language rather than gradients, producing executor-specific skills for compact models. Across five benchmarks on Qwen3.5-9B and Qwen3.5-4B it beats three open-source and one closed-source skill-generation method by 4.3 to 20.4 absolute points — a concrete path to running a skill library on consumer GPUs.