# ELF-REG

The work scales Embedded Language Flows (ELF) to mathematical reasoning and code generation and introduces ELF-REG, which uses a frozen autoregressive teacher to supervise intermediate denoiser features and to supply a global representation jointly denoised with the response (REPA+REG); evaluated on GSM8K, MATH-500, HumanEval, and MBPP, ELF-REG-L reaches 55.96% pass@1 on GSM8K at 64 NFE and 13.39% on MATH-500 and 22.56% on HumanEval at 128 NFE, outperforming the evaluated comparable-scale dLMs on GSM8K and code and improving MATH-500 from the ELF-L baseline of 10.55% to 13.39%, while the same task-specific checkpoints support strong low-NFE performance through early-stop without few-step training, reaching 41.21% HumanEval pass@10 at 16 NFE.

The work scales Embedded Language Flows (ELF) to mathematical reasoning and code generation and introduces ELF-REG, which uses a frozen autoregressive teacher to supervise intermediate denoiser features and to supply a global representation jointly denoised with the response (REPA+REG); evaluated on GSM8K, MATH-500, HumanEval, and MBPP, ELF-REG-L reaches 55.96% pass@1 on GSM8K at 64 NFE and 13.39% on MATH-500 and 22.56% on HumanEval at 128 NFE, outperforming the evaluated comparable-scale dLMs on GSM8K and code and improving MATH-500 from the ELF-L baseline of 10.55% to 13.39%, while the same task-specific checkpoints support strong low-NFE performance through early-stop without few-step training, reaching 41.21% HumanEval pass@10 at 16 NFE.