Search papers, labs, and topics across Lattice.
This paper introduces SelFusion, a self-distillation framework designed to enhance the performance of diffusion language models (DLMs) by addressing the limitations of conventional knowledge distillation (KD). By employing two forward passes with varying masking probabilities and implementing bidirectional KD based on token-level correctness, SelFusion achieves significant improvements in generation quality. Experimental results demonstrate that models trained with SelFusion can outperform traditional LLM teachers, offering a viable solution to the latency and quality challenges faced by DLMs in practical applications.
SelFusion enables diffusion language models to surpass traditional autoregressive models in generation quality by leveraging a novel self-distillation approach.
Diffusion language models (DLMs) alleviate the inherent latency bottleneck of autoregressive (AR) large language models (LLMs), but their degraded generation quality limits practical applicability. Although knowledge distillation (KD) can be a promising direction for improving performance, we empirically find that naively applying conventional KD yields only marginal gains, or even degrades generation quality. Based on these observations, we propose a novel self-distillation framework for DLMs, namely SelFusion. To enable effective KD without an external teacher model, SelFusion performs two forward passes with different masking levels, defining the hard mode with a larger masking probability and the easy mode with a smaller masking probability. However, the easy mode is not always more accurate than the hard mode and can be overconfident on incorrect tokens. Thus, we introduce bidirectional KD between the two modes, which can dynamically determine the distillation direction based on token-level correctness. Experimental results on instruction-following tasks show that the proposed self-distillation substantially outperforms other KD methods with external LLM and DLM teachers. In many configurations, the student trained with SelFusion even surpasses the performance of the LLM teacher, providing a practical path toward improving DLM generation quality. Source code can be found at https://github.com/scai-research/SelFusion_official
