Alibaba's HDPO cuts AI agent tool overuse from 98% to 2%

Alibaba researchers introduced Hierarchical Decoupled Policy Optimization (HDPO) to train AI agents with separate accuracy and efficiency objectives.

Their multimodal Metis model cut redundant tool invocations from 98% to 2%, showing far less tool overuse.

Metis also achieved state-of-the-art reasoning results on key benchmarks, suggesting the method can improve both performance and deployment efficiency.

The work targets a major AI-agent pain point: deciding when to rely on tools versus internal knowledge to reduce latency, cost, and failures.