Metis：教会多模态智能体"三思而后行"——HDPO框架破解工具滥用难题

Hong Kong Chinese University research team proposes the HDPO (Hierarchical Decoupled Policy Optimization) framework to address the tool overuse problem in multimodal agents. The Metis model trained with HDPO maintains high accuracy while reducing tool calls by several orders of magnitude, opening a new path for efficiency optimization of multimodal agents. This work aims to teach agents "think twice before acting" and develop metacognitive abilities.

Multimodal agents with visual understanding can actively call external tools (search engines, calculators, APIs) but often overuse them—even for problems solvable with visual info alone. This brings two costs: frequent API calls cause delays, and redundant external info becomes noise. For example, an agent might call OCR then calculator for a simple image text calculation instead of doing it directly.

Existing RL solutions use scalar penalties for tool calls, but face a dilemma: too strong → agents avoid tools even when needed (task failure); too weak → efficiency signals are drowned by accuracy reward variance. Traditional coupled rewards make accuracy and efficiency compete, hard to optimize both.

HDPO framework decouples the two goals into orthogonal channels: 1) Accuracy channel: maximize task correctness without considering tool cost; 2) Efficiency channel: optimize tool use only on accurate trajectories via conditional advantage estimation. This "learn to walk then run" approach first builds task-solving ability then optimizes efficiency, simulating human metacognition.

Unlike traditional advantage estimation, conditional advantage estimation only uses successful trajectories. It compares tool efficiency among accurate paths—trajectories with fewer tool calls get positive efficiency signals, ensuring efficiency gains don't sacrifice accuracy.

Metis model based on HDPO was evaluated on multiple multimodal benchmarks. It maintained or improved accuracy while reducing tool calls by several orders of magnitude. This shows agents can develop metacognition: judging whether to use tools or solve problems independently.

For enterprises, HDPO reduces API costs (major operational expense), shortens response time, and enhances system stability (less reliance on external services). The work proves efficiency and ability aren't对立—"smart and thrifty" agents are more practical.

Metis code is open-sourced on GitHub. Future extensions: apply HDPO to other metacognitive abilities (time management, planning) to build general AI systems with self-monitoring and regulation.