AgentPatch Repairs Merged Agentic MLLMs for Weak Task Recovery

Merging specialized agentic multimodal large language models (MLLMs) into a single generalist model presents significant challenges. Key issues include the uneven preservation of capabilities with varying interaction complexity, leading to "weak tasks," and the forgetting of critical behaviors essential for long-horizon execution. This paper introduces AgentPatch, a novel, training-free framework designed to repair these merged MLLMs. AgentPatch employs a coarse-to-fine approach. It first selects a stable merged backbone, then restores diluted weak-task-specific signals through a process called Weak-Task Unique Residual Recovery. Finally, it applies an Agent-Guided Behavior-Critical Patch to recover decisive actions while explicitly protecting existing capabilities. This framework produces a single, static checkpoint without needing complex routing or ensembles. Experiments across multiple agentic and multimodal benchmarks demonstrate that AgentPatch effectively improves diverse merged backbones, mitigates weak-task degradation, and balances recovery with the preservation of complementary capabilities.