base_model:
- GSAI-ML/LLaDA-8B-Instruct
datasets:
- Jiayi-Pan/Countdown-Tasks-3to4
license: mit
pipeline_tag: text-generation
library_name: transformers
ESPO-Countdown-LLaDA-8B-Instruct-LoRA
This repository contains a LoRA adapter for the GSAI-ML/LLaDA-8B-Instruct model, fine-tuned on the Countdown task using the ELBO-based Sequence-level Policy Optimization (ESPO) framework, as described in the paper Principled RL for Diffusion LLMs Emerges from a Sequence-Level Perspective.
ESPO is a principled reinforcement learning framework designed for diffusion large language models (dLLMs). It addresses the fundamental challenges of adapting RL methods to dLLMs by treating entire sequence generation as a single action and leveraging the ELBO (Evidence Lower Bound) as a tractable sequence-level likelihood proxy. This approach resolves the mismatch between RL and the non-autoregressive nature of dLLMs, leading to significant performance improvements on mathematical reasoning, coding, and planning tasks.
- 📚 Paper: Principled RL for Diffusion LLMs Emerges from a Sequence-Level Perspective
- 🌐 Project Page: ESPO Demo
- 💻 Code: ML-GSAI/ESPO
Sample Usage
You can load and use this ESPO-fine-tuned LoRA adapter on top of the base LLaDA-8B-Instruct model. Below is a quick start example demonstrating how to perform inference.
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
from eval.generate_utils import generate # Note: `eval.generate_utils` is a custom module from the GitHub repository and needs to be accessible.
base_model_path = 'GSAI-ML/LLaDA-8B-Instruct'
peft_model_path = 'GSAI-ML/ESPO-Math' # Replace with 'GSAI-ML/ESPO-Countdown' for this specific model
tokenizer = AutoTokenizer.from_pretrained(base_model_path)
model = AutoModelForCausalLM.from_pretrained(
base_model_path, trust_remote_code=True,torch_dtype="bfloat16", device_map="cuda")
peft_model = PeftModel.from_pretrained(model, peft_model_path, device_map="cuda")
prompt = "The point $(0,0)$ is reflected over the vertical line $x=1$. When its image is then reflected over the line $y=2$, what is the resulting point?\
\
Write your answer in the form $(x, y)$ where $x$ and $y$ are real numbers."
messages = [{"role": "user", "content": prompt}]
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
output_ids = generate(peft_model, input_ids,tokenizer, steps=128, gen_length=256, temperature=0.9,remasking="low_confidence",)
output_text = tokenizer.batch_decode(output_ids[:, input_ids.shape[1]:], skip_special_tokens=True)[0]
print(output_text)
Citation
If you find ESPO useful in your research, please consider citing our paper:
@article{ou2025principledrldiffusionllms,
title={Principled RL for Diffusion LLMs Emerges from a Sequence-Level Perspective},
author={Jingyang Ou and Jiaqi Han and Minkai Xu and Shaoxuan Xu and Jianwen Xie and Stefano Ermon and Yi Wu and Chongxuan Li},
journal={arXiv preprint arXiv:2512.03759},
year={2025},
}