Towards Scalable RLVR: Multimodal Instruction Following Data Synthesis and Distillation
MIFS mitigates the generalization trade-offs typical of SFT, preserving core visual capabilities while significantly boosting instruction-following precision, and introduces a generative constraint protocol to synthesize diverse raw samples.
Yi-Rong Zeng, Sai Zhang, Yu-Xian Wang et al.
· 0 citations