CoCoDA: Co-evolving Compositional DAG for Tool-Augmented Agents

Short summary

CoCoDA introduces a compositional code DAG framework that enables language models to efficiently manage and retrieve from growing tool libraries without exceeding context budgets. It co-evolves the tool library with the planner, using typed signatures and behavioral specifications to filter candidates intelligently. An 8B student model achieves parity with a 32B teacher on mathematical reasoning and coding benchmarks.

• •Organizes tools in a directed acyclic graph (DAG) structure to reduce retrieval costs
• •Co-evolves tool library and planner together as new composite tools emerge
• •8B model matches 32B baseline on GSM8K and MATH benchmarks

Generated with AI, which can make mistakes.