license: apache-2.0
library_name: peft
base_model: GSAI-ML/LLaDA-8B-Instruct
pipeline_tag: text-generation
tags:
- reinforcement-learning
- diffusion-llm
- block-r1
Block-R1
This repository contains model checkpoints (LoRA adapters) for Block-R1, a benchmark for multi-domain reinforcement learning with block-based diffusion large language models (dLLMs).
Description
Block-R1 is designed to enhance block-based reasoning generation in diffusion LLMs. It investigates the role of block size from a domain conflict perspective during reinforcement learning (RL) post-training. The benchmark covers diverse domains including code, mathematics, puzzles, and general knowledge.
Key components include:
- Block-R1-41K Dataset: A dataset constructed with optimized training block sizes for multi-domain RL.
- b1 Method: A dynamic-size reasoning block method for dLLMs.
- RL Framework: Support for multiple RL algorithms for diffusion models such as Diffusion-GRPO, WD1, GDPO, and more.
Resources
- Paper: Block-R1: Rethinking the Role of Block Size in Multi-domain Reinforcement Learning for Diffusion Large Language Models
- Code: GitHub Repository
- Dataset: Block-R1 Dataset
Model Information
These weights are LoRA adapters trained on top of the LLaDA-8B-Instruct backbone. For detailed usage, training, and evaluation scripts, please refer to the official repository.
Citation
If you use this benchmark or the associated methods, please cite the following work:
@article{jiang2026breakblock,
title={{Break the Block: Dynamic-size Reasoning Blocks for Diffusion Large Language Models via Monotonic Entropy Descent with Reinforcement Learning}},
author={Jiang, Yan and Qiu, Ruihong and Huang, Zi},
journal={arXiv preprint arXiv:2605.02263},
year={2026}
}