Reinforced Agent Merging: Preserving Specialized Behaviors in Agentic Models
Автор: AI Paper Review
Загружено: 2026-01-23
Просмотров: 20
A new model merging technique called *RAM (Reinforced Agent Merging)* is proposed to solve the performance degradation problem that occurs when integrating agent models trained with reinforcement learning (RL). The existing merging method is optimized for the mapping fine-tuning (SFT) environment, so there is a limit to diluting the core signal in the process of processing scarce and unbalanced parameter updates unique to the RL model. RAM separates updated parameters into shared and unique areas, averages the shared area, and selectively preserves and rebalances the unique area to maintain the expertise of each model. As a result of the experiment, this method performed better than the existing method in various fields such as coding, tool use, and long-term memory, and succeeded in implementing an integrated general-purpose model with superior capabilities than individual professional models. As a result, this paper demonstrates the importance of distribution-aware merge strategies for efficient coupling of RL-based agents.
https://arxiv.org/pdf/2601.13572
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
