Dr. Karen Ullrich
@karen_ullrich
Followers
5K
Following
587
Media
37
Statuses
276
Research scientist at FAIR NY + collab w/ Vector Institute. ❤️ Machine Learning + Information Theory. Previously, PhD at UoAmsterdam, intern at DeepMind + MSRC.
she / her
Joined December 2013
#Tokenization is undeniably a key player in the success story of #LLMs but we poorly understand why. I want to highlight progress we made in understanding the role of tokenization, developing the core incidents and mitigating its problems. 🧵👇
15
95
606
RL has led to amazing advances in reasoning domains with LLMs. But why has it been so successful, and why does the length of the response increases during RL? In new work, we introduce a framework to provide conceptual and theoretical answers to these questions.
2
12
56
eBioMedicine (part of “The Lancet Discovery Science”) published results from Intensity Therapeutics' Phase 1/2 IT-01 study of INT230-6 in metastatic or refractory cancers, showing a 75% disease control rate & 11.9-month median overall survival. Nasdaq: INTS
0
0
13
One can manipulate LLM rankings to put any model in the lead—only by modifying the single character separating demonstration examples. Learn more in our new paper https://t.co/D8CzSpPxMU w/ Jingtong Su, Jianyu Zhang, @karen_ullrich , and Léon Bottou. 1/3 🧵
1
3
11
Y’all, I am at #COLM this week, very excited to learn, and meet old and new friends. Please reach out on Whova!
0
1
6
From the government shutdown to views on the state of our political discourse, @brentbuc and @ChrisLaneMA cover the latest data from our National Voter Trends (NVT) poll. 🧵 on turbulence, turnover, and taking sides in America today...
1
11
29
Check out the full paper here: https://t.co/fKicf2Rfha 🎓 Work by Jingtong Su, @KempeLab, @NYUDataScience, @AIatMeta
arxiv.org
Transformers have achieved state-of-the-art performance across language and vision tasks. This success drives the imperative to interpret their internal mechanisms with the dual goals of enhancing...
0
0
3
Plus, we generate importance maps showing where in the transformer the concept is encoded — providing interpretable insights into model internals.
1
0
3
SAMI: Diminishes or amplifies these modules to control the concept's influence With SAMI, we can scale the importance of these modules — either amplifying or suppressing specific concepts.
1
0
2
SAMD: Finds the attention heads most correlated with a concept Using SAMD, we find that only a few attention heads are crucial for a wide range of concepts—confirming the sparse, modular nature of knowledge in transformers.
1
0
2
How would you make an LLM "forget" the concept of dog — or any other arbitrary concept? 🐶❓ We introduce SAMD & SAMI — a novel, concept-agnostic approach to identify and manipulate attention modules in transformers.
3
13
78
Aligned Multi-Objective Optimization (A-🐮) has been accepted at #ICML2025! 🎉 We explore optimization scenarios where objectives align rather than conflict, introducing new scalable algorithms with theoretical guarantees. #MachineLearning #AIResearch #Optimization #MLCommunity
3
12
88
Our work got accepted to #ICLR2025 @iclr_conf! Learn more about tokenization bias and how to convert your tokenized LLM to byte-level LLM without training! See you in Singapore! Check out the code here:
github.com
Example implementation of "Exact Byte-Level Probabilities from Tokenized Language Models for FIM-Tasks and Model Ensembles" by Buu Phan, Brandon Amos, Itai Gat, Marton Havasi, Mat...
🎉Our paper just got accepted to #ICLR2025! 🎉 Byte-level LLMs without training and guaranteed performance? Curious how? Dive into our work! 📚✨ Paper: https://t.co/SCNSWtkB3G Github: https://t.co/rxUMkVfW8U...
3
4
28
0
0
4
🎉Our paper just got accepted to #ICLR2025! 🎉 Byte-level LLMs without training and guaranteed performance? Curious how? Dive into our work! 📚✨ Paper: https://t.co/SCNSWtkB3G Github: https://t.co/rxUMkVfW8U...
2
14
111
📢 My team at Meta is hiring visiting PhD students from CMU, UW, Berkeley, and NYU! We study core ML, optimization, amortization, transport, flows, and control for modeling and interacting with complex systems. Please apply here and message me: https://t.co/QvZI94hhyy
9
42
277
Excited to release EvalGIM, an easy-to-use evaluation library for generative image models. EvalGIM ("EvalGym") unifies metrics, datasets, & visualizations, is customizable & extensible to new benchmarks, & provides actionable insights. Check it out!
github.com
🦾 EvalGIM (pronounced as "EvalGym") is an evaluation library for generative image models. It enables easy-to-use, reproducible automatic evaluations of text-to-image models and su...
2
15
93
Scenes from the most haunted houses in America. Forget ghosts — it’s the smart devices that have been haunting you all along. From fridges to vacuums, they’re quietly collecting your data and selling it to the highest bidder.
7
18
110
Thursday is busy: 9-11am I will be at the Meta AI Booth 12.30-2pm Mission Impossible: A Statistical Perspective on Jailbreaking LLMs ( https://t.co/14dqRGaHJJ) OR End-To-End Causal Effect Estimation from Unstructured Natural Language Data ( https://t.co/29sGvMX8Ww)
0
0
8
For those into jailbreaking LLMs: our poster "Mission Impossible" today shows the fundamental limits of LLM alignment - and improved ways to go about it, nonetheless. With @karen_ullrich & Jingtong Su #2302 11am - 2pm Poster Session 3 East @NYUDataScience @AIatMeta #NeurIPS2024
2
4
37
Starting with Fei-Fei Li’s talk 2.30, after that I will mostly be meeting people and wonder the poster sessions.
0
0
3
Folks, I am posting my NeurIPS schedule daily in hopes to see folks, thanks @tkipf the idea ;) 11-12.30 WiML round tables 1.30-4 Beyond Decoding, Tutorial
0
0
10
Style built for the spotlight. Crafted for performance. Blake Snell wears JAXXON.
0
0
9