Jifan Zhang
@jifan_zhang
Followers
329
Following
306
Media
17
Statuses
190
Research Fellow @AnthropicAI | Ph.D. @WisconsinCS @WIDiscovery | Previously BS/MS @uwcse, @Meta @Google @Amazon
Joined April 2017
RT @AnthropicAI: Today we're releasing Claude Opus 4.1, an upgrade to Claude Opus 4 on agentic tasks, real-world coding, and reasoning. htt….
0
1K
0
Just finished compiling eight papers into my PhD thesis with Claude’s help. I feel lucky compared to the many PhDs who graduated before me. I am also extremely jealous of the many PhDs who will graduate after me.
1
0
14
RT @AnthropicAI: New Anthropic research: Persona vectors. Language models sometimes go haywire and slip into weird and unsettling personas….
0
924
0
RT @AnthropicAI: We’re running another round of the Anthropic Fellows program. If you're an engineer or researcher with a strong coding o….
0
187
0
RT @TmlrPub: Deep Active Learning in the Open World. Tian Xie, Jifan Zhang, Haoyue Bai, Robert D Nowak. Action editor: Vincent Fortuin. h….
openreview.net
Machine learning models deployed in open-world scenarios often encounter unfamiliar conditions and perform poorly in unanticipated situations. As AI systems advance and find application in...
0
1
0
When are we going to get AI agents managing experiments for us? Running large scale experiments always mess up my sleep🥲.
0
0
3
RT @FabienDRoger: Very cool result!.I would have not predicted that when the model inits are the same, distillation transmits so much hidde….
0
1
0
RT @saprmarks: Subliminal learning: training on model-generated data can transmit traits of that model, even if the data is unrelated. Thi….
0
22
0
RT @lyang36: 🚨 Olympiad math + AI:. We ran Google’s Gemini 2.5 Pro on the fresh IMO 2025 problems. With careful prompting and pipeline desi….
0
118
0
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!.
0
0
2
RT @stochasticlalit: It was amazing to be part of this effort. Huge shout out to the team, and all the incredible pre-training and post-tra….
deepmind.google
Our advanced model officially achieved a gold-medal level performance on problems from the International Mathematical Olympiad (IMO), the world’s most prestigious competition for young...
0
8
0
How far are LLMs away from making an entire set of IMO problems?.
1
0
2
RT @ajwagenmaker: How can we train a foundation model to internalize what it means to “explore”?. Come check out our work on “behavioral ex….
0
52
0
This work is driven by my first and amazing undergrad advisee Shyam Nuggehalli. Also in collaboration with @stochasticlalit and @rdnowak. An implementation of the algorithm is in the LabelBench repo:
github.com
Contribute to EfficientTraining/LabelBench development by creating an account on GitHub.
0
0
2
With anything that claims JUST WORKS, there's always an asterisk, but ours is a small one. If you have limited budget (less than 5-10 per class), this algorithm and active learning in general is not going to save you much. Your best bet will be some sort of diversity sampling.
1
0
2
The algorithm is also noise tolerant and supports batch labeling, a significant improvement over my previous algorithm GALAXY.
Have limited labeling budget for training neural networks and the underlying data is too unbalanced? Check out our ICML 2022 paper “GALAXY: Graph-based Active Learning at the Extreme”. Joint work w/ @JulianJKS and @rdnowak. (1/6).
1
0
1
The end result is amazing. On 30 different dataset settings, we see this algorithm consistently improve over uncertainty sampling and random sampling (as well as a suite of other more advanced algorithms). In fact, the more imbalanced your dataset is, the more our algorithm saves.
1
0
1
Our strategy simply labels around what we call the optimal separation threshold (OST), where the density of uncertain unlabeled examples roughly equalize. The actual algorithm in finding the OST is complicated, so you'll have to read the paper.
1
0
1