Compositional Skill Acquisition in Agentic Pipelines via Reinforcement Learning and Knowledge Distillation

We study compositional skill acquisition in small language models (SLMs) trained for agentic pipeline tasks via reinforcement learning with verifiable rewards (RLVR) and knowledge dis- tillation (KD). On a benchmark of 100 agentic coding tasks spanning four categories with vary- ing compositional structure, RLVR achieves near- perfect generalization on structured categories (error-recovery: 0.000→1.000; multi-file feature: 0.667→1.000) but scores 0.000 on sparse-reward categories where the base model rarely succeeds. KD on 100 tasks achieves the strongest overall result (+0.62 held-out improvement), uniquely solving the sparse-reward tool/API category that RLVR cannot. The two methods fail on comple- mentary categories for structurally distinct rea- sons, motivating a staged KD→RLVR training recipe.