arxivcs.AI2026-06-30
ReGRPO: Reflection-Augmented Policy Optimization for Tool-Using Agents
Tool-augmented vision-language models (VLMs) can solve multimodal, multi-step tasks by calling external tools, yet they remain fragile in practice. Existing works have two common gaps. Supervised fine-tuning (SFT) is built mostly on successful trajectories and offers little signa…