Researchers propose CoCoDA, a new framework that organizes tool libraries as compositional code DAGs (directed acyclic graphs) to help smaller language models efficiently use external skills without exceeding context budgets. The system enables an 8B parameter model to match or exceed performance of a 32B model on mathematical reasoning benchmarks like GSM8K and MATH by co-evolving the tool library and planner in tandem, using typed retrieval and composition-aware rewards.
Why it matters: As tool-augmented AI systems scale, managing library size and retrieval costs while maintaining reasoning quality is a critical efficiency problem—CoCoDA's approach could unlock practical deployment of smaller, cost-effective models for complex reasoning tasks.