SkillComposer's 4B Model Coaches a 27B AI to Outperform Itself
SkillComposer trains a 4B model to write, refine, and merge reusable agent skills at inference time, boosting a 27B executor by up to +4.5 points on agent benchmarks
Takeaways
- Core idea: SkillComposer trains a 4B model to create, merge, and refine reusable agent skills at inference time, without human-labeled examples.
- Key result: A 4B skill model improves a 27B executor by up to +4.5 on agent tasks and +3.4 on coding tasks across three benchmarks.
- Three operations: Create extracts skills from trajectories, Merge generalizes across similar skills, Improve refines skills from failure feedback.
- Three deployment modes: Offline (pre-built library), Online (real-time evolution, no prior data), and Hybrid (both combined).
- Cross-domain transfer: Training on agent tasks alone improved code benchmark scores by +1.5, confirming skill composition is a general-purpose meta-ability.
- Implementation: Built on Qwen3.5-4B with LlamaFactory and vLLM; no public code repo announced yet.
Getting an LLM agent to reliably complete multi-step tasks is hard enough. Getting it to learn from those tasks and carry that knowledge forward is harder still. A new paper introduces SkillComposer, a framework that teaches a small language model to autonomously create, refine, and combine reusable "skills" at inference time, without any ground-truth supervision. The result: a 4B model that can meaningfully improve the performance of a much larger 27B executor across both agent and coding tasks.
The skill transfer trap
The idea of equipping agents with reusable skills is not new. In the agent literature, skills are reusable procedures that an agent can call across tasks or states, and they matter because they turn a long task into shorter, stable subroutines, so an agent can generalize by selecting and composing what it already knows. The problem is in how those skills are built.
Current skill construction methods treat the problem as one-shot extraction, overlooking a fundamental tension: a skill tailored to the specific task fails to transfer, while the abstracted skill often provides insufficient guidance. In other words, you are stuck choosing between a skill that is too specific to reuse, or one so generic it barely helps. SkillComposer is designed to escape that trap.
Three operations, one lifecycle
Skill construction is decomposed into three learnable abilities: (1) Skill Create extracts reusable procedural knowledge from raw execution trajectories, (2) Skill Merge drives generalization by consolidating semantically similar skill pairs into broader, more transferable ones, and (3) Skill Improve drives specification by refining a skill to better capture task-specific patterns based on new execution experience. Together, these form a complete lifecycle for a skill: born from a trajectory, broadened by merging with peers, and sharpened by real-world feedback.