NVlabs/ToolOrchestra

Trains small orchestrator models via multi-turn reinforcement learning to coordinate diverse tools and specialist LLMs, optimizing jointly for outcome, efficiency, and preference rewards. Includes automatic synthetic data generation pipeline (ToolScale dataset) for scaling RL training across complex multi-step tasks. Integrates with web search, code interpreters, and multiple expert models (GPT-4o, Claude, Llama variants) through unified tool-calling interface, evaluated on GAIA, HLE, FRAMES, and τ²-Bench benchmarks.

677 stars. Actively maintained with 2 commits in the last 30 days.