Trajectory-Aware Verbalized Optimization for Multi-Agent Systems

ICLR 2026 Conference Submission14481 Authors

Keywords: Agentic Optimization, Agentic Evaluation, Multi-agent System, verbalized optimization

Abstract: Large language model (LLM)-based multi-agent systems have shown significant potential, but their effectiveness often depends on manually engineered prompts, which are refined through labor-intensive trial and error. While automatic optimization methods exist, they often rely on coarse, task-level outcomes, neglecting the rich trajectory-level information that captures how agents reason, coordinate, and fail. To address this gap, we propose a Trajectory-Aware Verbalized Optimization (TAVO) framework for prompt refinement in multi-agent systems. Inspired by reinforcement learning, TAVO introduces a credit assignment mechanism that decomposes interaction trajectories into sub-trajectories, linking specific reasoning and coordination steps to the final outcome. This generates fine-grained, process-level feedback. By modeling prompts as verbalized policies, TAVO translates this trajectory feedback into concrete editing instructions, which are aggregated across tasks for systematic refinement. Experiments on both collaborative and competitive multi-agent benchmarks demonstrate that our framework enhances system performance while reducing coordination costs, underscoring the value of leveraging trajectory-level signals to construct more adaptive and efficient LLM-based multi-agent systems.

Primary Area: generative models

Submission Number: 14481