llada-8b-dllm-registers-countdown-rl-r4
Chunked GRPO on Countdown starting from the mix60k R4 SFT main-triad checkpoint. Released at step 200, the step reported in tab:rl_results.
This checkpoint is part of the dLLM Registers project — register tokens as a bounded, trained, continuous channel for carrying decoding state across denoising windows in diffusion language models.
- Paper section: tab:rl_results (Countdown, registers arm)
- Carry channel: continuous register tokens (
channel_mode=registers,num_registers=4,tail_length=0) - Base model: GSAI-ML/LLaDA-8B-Base
- Training data: Countdown (3- and 4-number variants, on-policy GRPO rollouts)
- Training config:
diffu-grpo/amlt/mix60k_r4_countdown34_promptvisible_bridge_bptt_bonete54.yaml - Chunk size: C = 128 tokens
- Prompt dropout rate (Bernoulli per-trace CSG mask): 0.3
- RL task: Countdown
- GRPO updates: 200
- Initialized from:
albertge/llada-8b-dllm-registers-mix60k-r4 - Date uploaded: 2026-06-13
How to load
from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained("albertge/llada-8b-dllm-registers-countdown-rl-r4", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("albertge/llada-8b-dllm-registers-countdown-rl-r4", trust_remote_code=True)
To use the carry channel correctly at inference, see the evaluator at
eval/eval.py
and the wrapper eval/run_mix60k_full_eval.sh.
Key flags for this checkpoint:
--num_registers 4 --channel_mode registers --tail_length 0 --use_mask_token_for_registers
Repository
Training and eval code: https://github.com/lbertge/d1-registers
Citation
If you use this checkpoint, please cite the dLLM Registers paper:
@misc{dllm-registers-2026,
title = {Register Tokens for Unbounded Reasoning in Diffusion Language Models},
author = {Albert Ge and collaborators},
year = {2026},
note = {Preprint},
url = {https://github.com/lbertge/d1-registers}
}