Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
[ICLR 2026] Official code for TraceRL: Revolutionizing post-training for Diffusion LLMs, powering the SOTA TraDo series.
| Date | Stars |
|---|---|
| 2026-07-31 | 511 |
| 2026-08-05 | 516 |
| 2026-08-06 | 516 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
<div align="center">
<br>
<img src="assets/logo.png" width="200">
<h3>Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models</h3>
<h4>Most comprehensive framework for dLLM's and multimodal dLLM's post-training</h4>
</div>
<p align="center">
<a href="https://arxiv.org/abs/2509.06949">
<img
src="https://img.shields.io/badge/Paper-Arxiv-red?logo=arxiv&logoColor=red"
alt="CURE Paper on arXiv"
/>
<a href="https://huggingface.co/collections/Gen-Verse/trado-series-68beb6cd6a26c27cde9fe3af">
<img
src="https://img.shields.io/badge/Datasets-Hugging%20Face%20Data-orange?logo=huggingface&logoColor=yellow"
alt="Coding Datasets on Hugging Face"
/>
</a>
<a href="https://huggingface.co/collections/Gen-Verse/trado-series-68beb6cd6a26c27cde9fe3af">
<img
src="https://img.shields.io/badge/TraDo%204B/8B-Hugging%20Face%20Model-FFCC00?logo=huggingface&logoColor=yellow"
alt="ReasonFlux Coders on Hugging Face"
/>
</a>
<a href="https://yinjjiew.github.io/projects/dllmrl/">
<img
src="https://img.shields.io/badge/Blog-TraceRL-blue?logo=rss&logoColor=white"
alt="Blog"
/>
</a>
</p>
<p align="center">
<img src="assets/figure1.png" alt="Overview" width="750">
</p>
## 🌱 Features
- **Model Support**: [TraDo](https://arxiv.org/abs/2509.06949), [SDAR](https://github.com/JetAstra/SDAR), [Dream](https://github.com/DreamLM/Dream), [LLaDA](https://github.com/ML-GSAI/LLaDA), [MMaDA](https://github.com/Gen-Verse/MMaDA), [LLaDA-V](https://github.com/ML-GSAI/LLaDA-V), and [Diffu-Coder](https://github.com/apple/ml-diffucoder) Almost all open-sourced discrete diffusion language models are supported here.
- **Diverse Settings**: We support deployment, **SFT**, **RL** (with **optional value model** for variance reduction and **process reward model** for fine-grained supervision), and **RLHF** across diverse settings (**math, coding, multimodal**) and different architectures (**both full/block attention dLLMs**).
- **Inference Acceleration**: improved [KV-cache](https://github.com/NVlabs/Fast-dLLM/tree/main), [jetengine](https://github.com/Labman42/JetEngine/tree/0ddc55ad3fb712b6374515b78d656f420e1a7243) (based on nano-vllm), different sampling strategies, support multi-nodes, easy to build your own accelerated inference methods.
- **RL Training**: [TraceRL (support diffusion value model)](https://arxiv.org/abs/2509.06949), [coupled RL](https://github.com/apple/ml-diffucoder), [random masking RL](https://github.com/Gen-Verse/MMaDA), accelerated sampling, including Math, coding, and general RL tasks, support multi-nodes, easy to build your reinforcement learning methods across diverse settings
- **SFT**: [Block SFT](https://github.com/kuleshov-group/bd3lms), semi-AR SFT, random masking SFT, support multi-nodes and long-CoT finetune.
## 🧠 RL Methods (TraceRL) & Models (TraDo)
We propose **TraceRL**, a trajectory-aware reinforcement learning method for diffusion language models, which demonstrates the best performance among RL approaches for DLMs. We also introduce a diffusion-based value model that reduces variance and improves stability during optimization.
<p align="center">
<img src="assets/sft.png" width="48%"/>
<img src="assets/rl.png" width="48%"/>
</p>
Based on TraceRL, we derive a series of diffusion language models, **TraDo**, which achieve state-of-the-art performance on math and coding reasoning tasks. TraDo-4B-Instruct and TraDo-8B-Instruct are trained solely with TraceRL, while the first long-CoT diffusion language model, TraDo-8B-Thinking, is obtained through a combination of TraceRL and long-CoT data SFT. TraDo models challenge AR models with strong empirical results, as shown in the following table.
<p align="center">
<img src="assets/maintable.png" alt="Main Table" width="750">
</p>
We can download and try our model:
```python
from transformers import AutoModelForCausalLM, AutoTokeniExcerpt of 14,297 characters
Read on GitHub
256
Ling Yang
13
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:8f8746bfe9822977, topic:rlhf
matched fp:8f8746bfe9822977, topic:large-language-models
matched fp:8f8746bfe9822977, topic:code-generation