Dawei Li
@Dawei_Li_ASU
Followers
395
Following
463
Media
32
Statuses
131
CS PhD @ ASU https://t.co/JQ2VAX1cZM LLMs, NLP, Data Mining Founder of Oracle-LLM: https://t.co/BJtkt1Bt8i
Joined November 2024
🏆 Best Paper Award at DIG-BUG@ICML 2025! 📢📢Thrilled to share that our work "Preference Leakage: A Contamination Problem in LLM-as-a-Judge (
https://t.co/NNOw7tvd6l)" has received the Best Paper Award at the #ICML2025 Workshop on Data in Generative Models (DIG-BUG)! This is my
0
5
18
📢 NeurIPS-W Paper — Compression × Safety 📢 What happens when you compress an LLM… and accidentally make it easier to jailbreak? 🤔 Excited to share our new work: “Compressed but Compromised: A Study of Jailbreaking in Compressed LLMs.”
1
7
15
Not able to attend #NeurIPS2025 in person, here are two papers from #DMML and feel free to reach out to our other lab members at the conference! @ChengshuaiZhao, @ujun_asu, Bohan Jiang and Zhen Tan
0
1
12
🧠 New work on "honest" deductive reasoning in LLMs: Can models learn when they don't have enough information to answer? 😆 Join us this Saturday at 10 PM EST for a talk by @Jiarui_Liu_ on Honest large language models! 📷 Register here: https://t.co/EEZ5eYKo91 ⏲️ 10 PM EST on
1
7
15
Interpretability meets Agents! 🤖 Introducing SAGE: An Agentic Explainer Framework for Interpreting SAE Features. Instead of passive, single-pass explanations, SAGE acts like a scientist. It uses an iterative loop to "debug" feature meanings. #Interpretability #MechInterp #SAE
3
4
16
Can AI agents autonomously explore, synthesize, and discover knowledge like researchers? 🤖🔬 Introducing a comprehensive survey on Deep Research (DR) systems, where LLMs evolve from passive text generators into autonomous agents capable of long-horizon reasoning and verifiable
10
58
225
A bug in #OpenReview just exposed every reviewer’s identity across all conferences.😂 ⚔️Next week’s in-person NeurIPS be like: 👇
2
23
164
Happy Thanksgiving!! 🦃 👏 Join us for a talk by Huanxin Sheng on Analyzing Uncertainty of LLM-as-a-Judge on this Saturday, 11/29@9 pm EST! ⏰ Event Registration here: https://t.co/cRAbAzVBqB Huanxin is a second-year PhD student in Computer Science at the University of
0
6
10
Before the GenAI wave, distillation usually targeted small output spaces, like transferring label logits in classification tasks. But today’s foundation models focus more on scaling and learning intermediate reasoning steps, from text CoT to multi-modal deep thinking. When we try
Most people assume you need a massive dataset to distill flow models. We challenge that. Is data actually necessary? Or perhaps it is a liability? Introducing FreeFlow: We achieve SOTA (1.49 FID on ImageNet-512) 1-step image generation without a single data sample. 🧵👇[1/n]
0
13
82
Review your paper using AI before the reviewers use it for you🥲
Releasing a new "Agentic Reviewer" for research papers. I started coding this as a weekend project, and @jyx_su made it much better. I was inspired by a student who had a paper rejected 6 times over 3 years. Their feedback loop -- waiting ~6 months for feedback each time -- was
0
0
8
Join the talk and learn more about trustworthy foundation model building👉
👏👏Join us for a talk by Yuji Zhang @Yuji_Zhang_NLP on Developing Robust and Trustworthy Foundation Models on this Saturday@9 pm! ⏲️Event Registration here: https://t.co/NMi3l5Fqtx Yuji is a postdoctoral researcher at UIUC working on robust and trustworthy foundation models,
0
0
4
After half a year since #Deepseek's GRM and other reasoning RM have been proposed, it's good to see thinking-LLM-as-a-judge has been scaled up to commercial LLMs! 👉What to learn more about the opportunities and challenges of LLM-as-a-judge? Check our website:
llm-as-a-judge.github.io
LLM-as-a-judge.
Post-training at xAI: Over the past few months, our team of a dozen overhauled the RL recipe using user preference on real conversations; and agentic reward models that grade using strong reasoning capabilities. We also scaled up RL an order of magnitude more than the existing
0
0
6
With ~21% of #ICLR reviews flagged as AI-generated, maybe it’s time to ask: who’s judging your paper — a human or an LLM? In our recent paper: "Who’s Your Judge? On the Detectability of LLM-Generated Judgments" we study how to detect LLM-generated judgments and reviews — even
ICLR authors, want to check if your reviews are likely AI generated? ICLR reviewers, want to check if your paper is likely AI generated? Here are AI detection results for every ICLR paper and review from @pangramlabs! It seems that ~21% of reviews may be AI?
0
1
14
With OpenAI GPT-5.1 and Baidu ERNIE 5.0 launching the same day, AI has entered the era of native multimodality + agentic reasoning. The frontier is no longer larger context windows: it’s richer context worlds. ERNIE 5.0 isn’t just a product milestone, it’s a research invitation
0
0
0
ERNIE 5.0 Preview ranked #2 globally on LMArena (text), beating many closed models in reasoning and creative writing. But its significance is deeper: it offers a new testbed for grounded semantics, cross-modal alignment, and emotion understanding in one model.
1
0
0
Built on PaddlePaddle, ERNIE 5.0 trains 2.4 trillion parameters with <3 % active (super-sparse MoE). It uses asynchronous multimodal encoder separation, FP8 mixed precision, and adaptive off-loading — innovations that make trillion-scale multimodal training practical, not
1
0
0
The Platonic Representation Hypothesis ( https://t.co/fvou7DnLDe) proposes that as models scale, their internal representations of the world converge — regardless of architecture or modality — toward a shared “Platonic” structure of reality. This idea reframes multimodality: if
1
0
13
Join us this Saturday at 10 PM EST for a talk by Yueqi Song @yueqi_song on Agent Data Protocol! 🤖 🎟️ Register here: https://t.co/DG5TjQxgAV Learn about a comprehensive framework that enables: ✨ Reproducible research in agent behaviors ✨ Seamless dataset integration ✨ 1.3M
0
13
33
📢CIKM 2025 Tutorial: Generative AI & Synthetic Data for Data Mining 💡Join us at CIKM 2025 on November 10 (9:00 AM, Seoul time) for our tutorial — “Generative Models for Synthetic Data: Transforming Data Mining in the GenAI Era.” ( https://t.co/i8i6ypxwd0) We’ll explore how
0
0
8
I am heading to #EMNLP2025 and will present two papers about LLM-based assessment. Feel free to connect with me if you are interested in data synthesis, LLM-as-a-judge or any other topics related to machine learning, data mining and LLMs!
2
0
14
Nice blog! It shares similar points with our previous DRP paper ( https://t.co/bKOMgQ1tUC) that leverages a strong teacher to revise and polish the student model's reasoning trajectory rather than direct distillation for better learnability.
Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other
0
0
6