ELF-REG Scales Continuous Diffusion Language Models to Reasoning Tasks
September 25, 2026
ELF-REG improves continuous diffusion language models using representation alignment and entanglement with a frozen autoregressive teacher. The model achieves 55.96% pass@1 on GSM8K and 22.56% on HumanEval using 128 network function evaluations.
HOW THIS AFFECTS YOU
●
builderYou may soon use diffusion-based decoders to generate all tokens in parallel for faster inference.
●
researcherYou can explore new training objectives for non-autoregressive modeling via representation alignment.