blanchon/elf-reg-demo-weights

ELF-REG demo weights (EMA-0.9999, trimmed)

EMA-0.9999 inference weights for ELF-REG: Scaling Continuous Diffusion Language Models to Reasoning Tasks (arXiv:2609.29102), by Zeyu Michael Li, William Xingxu Chen, Bingshuo Qian, Jiayin Liu and Xiang Cheng (Duke University / Tsinghua University).

These are trimmed mirrors of the official checkpoints published by the first author at zl310/elf-reg — same directory layout, same config.yml files. The originals contain the optimizer state and multiple EMA copies (39 GB for the four demo checkpoints); these files keep only ema_params1["0.9999"], the exact weights the paper's headline evaluation loads (10.7 GB total). Content of the EMA state dicts is byte-identical to the originals; nothing was retrained or re-quantized (fp32).

Official code (MIT license): lizeyu090312/scaling_dLM, built on lillian039/ELF (arXiv:2605.10938).

Task	Checkpoint	Model	Size
GSM8K	gsm8k/elf_b_repa_reg/checkpoint_56352 (epoch 12)	ELF-REG-B (104M+156M)	1.0 GB
GSM8K	gsm8k/elf_l_repa_reg/checkpoint_56352 (epoch 12)	ELF-REG-L (652M+157M)	3.2 GB
MATH-500	math500/elf_l_repa_reg/checkpoint_55824 (epoch 8)	ELF-REG-L	3.2 GB
Code	code/elf_l_repa_reg/checkpoint_64248 (epoch 12)	ELF-REG-L	3.2 GB

License

The upstream code is MIT-licensed. The upstream checkpoint weights carry no explicit license; they are mirrored here without modification for use by the demo Space blanchon/elf-reg-demo, with attribution to the authors. If you are an author and would like this mirror changed or removed, please open a discussion.

Verification

GSM8K pass@1 at NFE 64 (early-stop ρ=8, SCCFG 3, EMA 0.9999), official eval code and protocol:

Source	pass@1
Paper (Table 2, 16 seeds, full test set)	38.11% ± 1.07
This stack (subset: first 160 questions, 4 seeds)	40.00% ± 0.51
This stack (full 1319-question test set, 4 seeds)	38.40% ± 0.61

Citation

@misc{elf_reg2026,
      title={ELF-REG: Scaling Continuous Diffusion Language Models to Reasoning Tasks},
      author={Zeyu Michael Li and William Xingxu Chen and Bingshuo Qian and Jiayin Liu and Xiang Cheng},
      year={2026},
      eprint={2609.29102},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.29102},
}