MRC-MMAgent: Reinforcement Learning for Multi-Role Collaboration in Multimodal Tool-Using Agents
Multi-role RL lets a multimodal agent inspect evidence and resume when support is weak. It trains on the original baseline dataset only.
The claim
Multi-role RL lets a multimodal agent inspect evidence and resume when support is weak.
The problem
Multimodal agents answer with tools as well as images.
RL pipelines mostly train one role with a terminal reward.
Humans inspect evidence, reformulate queries and revise conclusions while searching.
How it works
A collaborative loop joins solving, inspection and continuation.
The agent gathers evidence through tools.
It judges whether the last trajectory supports the answer.
If support is weak, it resumes problem solving.
Every role's decisions enter the same policy objective.
Figures
What changes
One objective scores task success, interaction validity and successful correction.
Full-trajectory policy optimization covers every decision produced by all three roles.
The model beats its base model on the original baseline training set.
Limits
The abstract reports no benchmark names and no numeric figures.
Training uses only the baseline dataset, with no newly collected data.
The gains appear only on the benchmarks the authors actually ran.
Source · arXiv