yjyddq/EOSER-ASS-RL

Official Repository of "Taming Masked Diffusion Language Models via Consistency Trajectory Reinforcement Learning with Fewer Decoding Step"

36
/ 100
Emerging

This project offers novel techniques to enhance Masked Diffusion Language Models (MDLMs). It introduces methods like EOS Early Rejection (EOSER) decoding and an Ascending Step-Size (ASS) scheduler, along with Consistency Trajectory Group Relative Policy Optimization (CJ-GRPO). These innovations aim to make MDLMs more efficient, allowing them to achieve competitive performance with fewer decoding steps. Developers and researchers working with diffusion models for natural language generation will find this useful for improving model speed and output quality.

Use this if you are a machine learning researcher or developer working with Masked Diffusion Language Models and need to reduce the number of decoding steps while maintaining or improving performance.

Not ideal if you are looking for a pre-trained model or an application-ready tool, as this project focuses on foundational algorithmic improvements for MDLMs.

natural-language-generation machine-learning-research diffusion-models language-model-optimization deep-learning-algorithms
No Package No Dependents
Maintenance 10 / 25
Adoption 7 / 25
Maturity 15 / 25
Community 4 / 25

How are scores calculated?

Stars

27

Forks

1

Language

Python

License

Apache-2.0

Last pushed

Mar 09, 2026

Commits (30d)

0

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/diffusion/yjyddq/EOSER-ASS-RL"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.