Jifan Zhang
@jifan_zhang
Research Fellow @AnthropicAI | Ph.D. @WisconsinCS @WIDiscovery | Previously BS/MS @uwcse, @Meta @Google @Amazon
Deep Active Learning in the Open World Tian Xie, Jifan Zhang, Haoyue Bai, Robert D Nowak. Action editor: Vincent Fortuin. openreview.net/forum?id=Hkmym… #classification #annotation #classes
When are we going to get AI agents managing experiments for us? Running large scale experiments always mess up my sleep🥲
Very cool result! I would have not predicted that when the model inits are the same, distillation transmits so much hidden information about the teacher. (This is much more powerful than emergent-misalignment-like phenomenon!)
New paper & surprising result. LLMs transmit traits to other models via hidden signals in data. Datasets consisting only of 3-digit numbers can transmit a love for owls, or evil tendencies. 🧵
Subliminal learning: training on model-generated data can transmit traits of that model, even if the data is unrelated. Think: "You can learn physics by watching Einstein do yoga" I'll discuss how this introduces a surprising pitfall for AI developers 🧵x.com/OwainEvans_UK/…
New paper & surprising result. LLMs transmit traits to other models via hidden signals in data. Datasets consisting only of 3-digit numbers can transmit a love for owls, or evil tendencies. 🧵
🚨 Olympiad math + AI: We ran Google’s Gemini 2.5 Pro on the fresh IMO 2025 problems. With careful prompting and pipeline design, it solved 5 out of 6 — remarkable for tasks demanding deep insight and creativity. The model could win gold! 🥇 #AI #Math #LLMs #IMO2025
Congrats Lalit and the GDM team for winning🏅!
It was amazing to be part of this effort. Huge shout out to the team, and all the incredible pre-training and post-training efforts that ensure Gemini is the leading frontier model! deepmind.google/discover/blog/…
It was amazing to be part of this effort. Huge shout out to the team, and all the incredible pre-training and post-training efforts that ensure Gemini is the leading frontier model! deepmind.google/discover/blog/…
How far are LLMs away from making an entire set of IMO problems?
How can we train a foundation model to internalize what it means to “explore”? Come check out our work on “behavioral exploration” at ICML25 to find out!
Looking forward to seeing folks tomorrow afternoon!
Working in active learning, I often hear a common complaint — active learning doesn’t work sometimes. Our ICML 2025 paper JUST WORKS* when you have class imbalance. @rdnowak will present it Tuesday afternoon. Short 🧵below Paper: arxiv.org/abs/2312.09196
xAI launched Grok 4 without any documentation of their safety testing. This is reckless and breaks with industry best practices followed by other major AI labs. If xAI is going to be a frontier AI developer, they should act like one. 🧵
Releasing HumorBench today. Grok 4 is🥇 on this uncontaminated, non-STEM humor reasoning benchmark. 🫡🫡@xai Here are couple things I find surprising👇 1. this benchmark yields an almost perfect rank correlation with ARC-AGI. Yet the task of reasoning about New Yorker style…
Whoa... Grok 4 beats o3 on our never-released benchmark: HumorBench, a non-STEM reasoning benchmark that measures humor comprehension. The task is simple: given a New Yorker Caption Contest cartoon and caption, explain the joke.