---
license: apache-2.0
language:
- en
base_model: GSAI-ML/LLaDA-8B-Base
library_name: transformers
pipeline_tag: text-generation
tags:
- diffusion-language-model
- register-tokens
- chunked-reasoning
- dllm-registers
- chunked-grpo
---

# llada-8b-dllm-registers-longarith-rl-t4

Chunked GRPO on LongArithmetic starting from the mix60k T4 SFT main-triad checkpoint. Released at step 400, 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 (LongArithmetic, Markovian Thinking arm)
- **Carry channel**: discrete Markovian-Thinking tail (`channel_mode=tail`, `num_registers=0`, `tail_length=4`)
- **Base model**: [GSAI-ML/LLaDA-8B-Base](https://huggingface.co/GSAI-ML/LLaDA-8B-Base)
- **Training data**: LongArithmetic (5-8 operations, add/sub20, on-policy GRPO rollouts)
- **Training config**: [`diffu-grpo/amlt/mix60k_t4_longarith_expr5678_state_bonete54.yaml`](https://github.com/lbertge/d1-registers/blob/main/diffu-grpo/amlt/mix60k_t4_longarith_expr5678_state_bonete54.yaml)
- **Chunk size**: C = 128 tokens
- **Prompt dropout rate** (Bernoulli per-trace CSG mask): 0.0
- **RL task**: LongArithmetic
- **GRPO updates**: 400
- **Initialized from**: [`albertge/llada-8b-dllm-registers-mix60k-t4`](https://huggingface.co/albertge/llada-8b-dllm-registers-mix60k-t4)
- **Date uploaded**: 2026-06-13

## How to load

```python
from transformers import AutoModel, AutoTokenizer
model = AutoModel.from_pretrained("albertge/llada-8b-dllm-registers-longarith-rl-t4", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("albertge/llada-8b-dllm-registers-longarith-rl-t4", trust_remote_code=True)
```

To use the carry channel correctly at inference, see the evaluator at
[`eval/eval.py`](https://github.com/lbertge/d1-registers/blob/main/eval/eval.py)
and the wrapper [`eval/run_mix60k_full_eval.sh`](https://github.com/lbertge/d1-registers/blob/main/eval/run_mix60k_full_eval.sh).
Key flags for this checkpoint:
`--num_registers 0 --channel_mode tail --tail_length 4`

## Repository

Training and eval code: [https://github.com/lbertge/d1-registers](https://github.com/lbertge/d1-registers)

## Citation

If you use this checkpoint, please cite the dLLM Registers paper:

```bibtex
@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}
}
```