Zhihui Xie
@_zhihuixie
Followers
372
Following
1K
Media
19
Statuses
169
Ph.D. student @hkunlp2020 | Intern @AIatMeta | Previously @sjtu1896
Joined July 2019
🚀 Thrilled to announce Dream-Coder 7B — the most powerful open diffusion code  LLM to date.
2
34
112
RT @FaZhou_998: Apart from the performance, it’s pure entertainment just watching Qwen3‑Coder build Qwen Code all by itself. Agentic coding….
0
10
0
RT @sivil_taram: 🔥 LLMs can fix bugs, but can they make your code faster? We put them to the test on real-world repositories, and the resul….
0
17
0
RT @RuiYang70669025: My coauthor @hc81Jeremy will present EmbodiedBench at ICML 2025! 🤖.Oral Session 6A.📍 West Hall C 🕧July 17 3:30-3:45 pm….
0
3
0
RT @gm8xx8: Follow-up to Dream 7B, now focused on code:. Dream-Coder 7B is a diffusion-based code LLM from HKU + Huawei Noah’s Ark, built o….
0
6
0
RT @JiachengYe15: 📢 Update: Announcing Dream's next-phase development. - Dream-Coder 7B: A fully open diffusion LLM for code delivering s….
0
22
0
For more details, please check out:
We present DreamOn: a simple yet effective method for variable-length generation in diffusion language models. Our approach boosts code infilling performance significantly and even catches up with oracle results.
0
0
2
🙌 This project couldn’t have happened without the amazing team: @JiachengYe15, @linzhengisme, @jiahuigao3, Jingwei Dong, @WilliamZR7, @xlzhao_hku, @sansa19739319, Xin Jiang, Zhenguo Li & @ikekong!.
0
0
10
We provide complete transparency throughout the development pipeline — data-processing scripts, implementation code, and model weights are all publicly available. 📑 Notion: 📗 Blog post: 🤗 Models: 💻.
huggingface.co
1
1
12
RT @WilliamZR7: We present DreamOn: a simple yet effective method for variable-length generation in diffusion language models. Our approac….
0
28
0
- SFT: We introduce random-truncation and PAD-penalty strategy to mitigate low sample efficiency and generation-length instability. - RLVR: We apply a tailored version of GRPO for diffusion LLMs to further refine the model through verifiable unit-test rewards.
1
0
8
Our training recipe includes:.- Adaptation: Dream-Coder 7B is adapted from Qwen2.5-Coder 7B base, achieving top-tier coding performance among open AR and diffusion LLMs while preserving strong language, math, and scientific-reasoning skills.
1
0
7
(2) For infilling, we introduce an variant DreamOn-7B, which dynamically adjusts masked spans during infilling, expanding or contracting them to fit the exact target length.
2
0
8
- Interleaved reasoning that echoes input & solves step-by-step
1
0
7
Appealing features.(1) Dream-Coder adaptively generates code in whichever order the task demands:.- Sketch-first scaffolds for stdin/stdout problems
1
0
7
Trained entirely on publicly available datasets, Dream-Coder 7B:. - Outperforms existing open-weight diffusion model on code tasks by a large margin;.- Achieves 21.4 % pass@1 on LiveCodeBench (2410-2505), matching or surpassing proprietary and open-weight AR models such as.
1
0
9
RT @RekaAILabs: 📢 We are open sourcing ⚡Reka Flash 3.1⚡ and 🗜️Reka Quant🗜️. Reka Flash 3.1 is a much improved version of Reka Flash 3 that….
0
49
0