Reinforcement Fine-Tuning Enables MLLMs Learning Novel Tasks Stably (2506.23508v1)

Abstract: Post-training algorithms such as Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT) are widely used to adapt multimodal LLMs to downstream tasks. While effective at task adaptation, their impact on prior knowledge remains unclear. In this paper, we introduce jigsaw puzzles as a novel task absent from existing pretraining corpora and systematically study the behavior of SFT and RFT on an open-source multimodal model, Qwen2.5-VL. Our experiments reveal a sharp trade-off: SFT enables rapid task acquisition but leads to catastrophic forgetting, whereas RFT learns more slowly on novel tasks but maintains prior knowledge. We analyze this phenomenon through the lens of learning dynamics, showing that RFT reinforces correct samples that are naturally aligned with the base model's probability landscape, mitigating interference with prior knowledge. Moreover, supervised training on correct RFT-simulated rollouts allows SFT to preserve knowledge while rapidly learning new tasks. These findings suggest that data distribution, rather than algorithmic differences, plays a central role in forgetting, and highlight RFT's potential for stable continual learning in multimodal LLMs.