ToolRLA: Fine-Grained Reward Decomposition for Tool-Integrated Reinforcement Learning Alignment in Domain-Specific Agents

Researchers developed ToolRLA, a three-stage reinforcement learning pipeline that significantly improves AI agents' ability to use external tools and APIs for domain-specific tasks. The system achieved 47% higher task completion rates and 93% lower regulatory violations when deployed in a real-world financial advisory copilot serving 80+ advisors with 1,200+ daily queries.

• →ToolRLA uses fine-grained reward decomposition across four dimensions: format validity, tool selection, efficiency, and regulatory compliance.
• →The system achieved 91% task completion rate compared to 62% with previous methods in production deployment.
• →Regulatory violations dropped from 12% to 0.8% through multiplicative reward composition and large compliance penalties.
• →The three-stage pipeline includes Supervised Fine-Tuning, Group Relative Policy Optimization, and Direct Preference Optimization.
• →Performance improvements were validated across multiple benchmarks including ToolBench and API-Bank datasets.