Zihao Li
@_Violet24K_
Followers
90
Following
123
Media
13
Statuses
85
Ph.D. candidate @siebelschool @UofIllinois | (ex-)intern @Amazon @MSFTResearch
Joined January 2023
Proud to share our own work here โ a simple diffusion recipe that turns standard BERTs into chat models. All code, checkpoints, and training curves are fully open in our repo. Really excited to push diffusion-based text generation further!
(1/n) ๐จ BERTs that chat: turn any BERT into a chatbot with diffusion hi @karpathy, we just trained a few BERTs to chat with diffusion โ we are releasing all the model checkpoints, training curves, and recipes! Hopefully this spares you the side quest into training nanochat with
0
2
5
With DRAMA being a new paradigm that unifies a wide range of sub-problems, DRAMA-Bot is just the beginning โ imagine complex data integration, cleaning, etc, all in one agent. Canโt wait to see what DRAMA unfolds next: DATA SCIENCE IS FULL OF DRAMA๐ญ๐ค
Real-world data science (especially in the social sciences) starts with collecting and structuring data from open domains, yet existing AI agents either assume access to ready-to-query databases or stop at surface-level retrieval and summarization. To augment data scientists,
3
1
8
๐ฅThrilled to share our new work Reinforce-Ada, which fixes signal collapse in GRPO ๐ฅณNo more blind oversampling or dead updates. Just sharper gradients, faster convergence, and stronger models. โ๏ธ One-line drop-in. Real gains. https://t.co/kJTeVek1S3
https://t.co/7qLywG2KWR
7
24
179
Benchmarks don't just measure AI; they define its trajectory. Today, thereโs a shortage of truly challenging and useful benchmarks for LLMs โ and we believe future forecasting is the next frontier. Introducing TradeBench. https://t.co/VI6nJMT58W A live-market benchmark where
2
6
35
๐ Flow Matching Meets Biology and Life Science: A Survey Flow matching is emerging as a powerful generative paradigm. We comprehensively review its foundations and applications across biology & life science๐งฌ ๐Paper: https://t.co/ynsegKOgXz ๐ปResource:
github.com
A curated list of resources for "Flow Matching Meets Biology and Life Science: A Survey" - Violet24K/Awesome-Flow-Matching-Meets-Biology
0
10
10
๐ง Letโs teach LLMs to learn smarter, not harder๐ฅ[ https://t.co/DMjQaWyceE] ๐คHow can LLMs verify complex scientific information efficiently? ๐We propose modular, reusable atomic reasoning skills that reduce LLMsโ cognitive load to verify scientific claims with little data.
7
36
107
๐ฒ Not only reasoning?! Inference scaling can now boost LLM safety! ๐ Introducing Saffron-1: - Reduces attack success rate from 66% to 17.5% - Uses only 59.7 TFLOP compute - Counters latest jailbreak attacks - No model finetuning On the AI2 Refusals benchmark. ๐ Paper:
2
21
74
Joining the Hundredaire Club๐ฏ as a junior member. A.M. Turing committee please call me anytime.
1
0
12
๐ค New preprint: We propose ten principles of AI agent economics, offering a framework to understand how AI agents make decisions, influence social interactions, and participate in the broader economy. ๐ Paper: https://t.co/kI1ze1WQL5
1
10
14
๐จ ICML โ25 SPOTLIGHT ๐จ Taming Knowledge Conflict in Language Models ๐ค Why does your LLM sometimes echo the prompt but other times rely on its โbuilt-inโ facts? ๐ญ Can we toggle between parametric memory and fresh context without fine-tuning? ๐ฌ Curious about LLM internals,
2
4
20
We'll present 4 papers and 1 keynote talk at #ICLR2025. Prof. Jingrui He and Prof. Hanghang Tong will be at the conference. Let's connect! โ๏ธ
0
3
8
๐จCall for Papers - ๐ ๐๐ผ๐-๐๐ฒ๐ป๐๐ @ ๐๐๐ ๐ฎ๐ฌ๐ฎ๐ฑ Join us at the Workshop on ๐ ๐ฎ๐ฐ๐ต๐ถ๐ป๐ฒ ๐๐ฒ๐ฎ๐ฟ๐ป๐ถ๐ป๐ด ๐ผ๐ป ๐๐ฟ๐ฎ๐ฝ๐ต๐ in the Era of ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐๐ co-located with #KDD2025! ๐Website: https://t.co/hpeImZm9KR ๐Submission Link: https://t.co/z7BXvjmoiB
0
8
18
๐ฌGraph Self-Supervised Learning Toolkit ๐ฅWe release PyG-SSL, offering a unified framework of 10+ self-supervised choices to pretrain your graph foundation models. ๐Paper: https://t.co/fwlWTsmquK ๐ปCode: https://t.co/oNaz18zCht Have fun!
github.com
Graph Self-Supervised Learning Toolkit. Contribute to iDEA-iSAIL-Lab-UIUC/pyg-ssl development by creating an account on GitHub.
0
2
3
๐New findings of knowledge overshadowing! Why do LLMs hallucinate over all true training data? ๐คCan we predict hallucinations even before model training or inference? ๐Check out our new preprint: [ https://t.co/Rzq7zFyzKF] The Law of Knowledge Overshadowing: Towards
6
34
123