Stanley Wei
@stanleyrwei
Followers
106
Following
24
Media
2
Statuses
19
PhD student @Princeton. Theoretical foundations of machine learning and LLMs. Previously CS + Math @UTAustin.
Joined July 2022
Our new (algorithmic) coding eval benchmark!. Fun collab with a large team of my competitive programming friends - we performed large scale manual annotation of contest problems to pin down exact areas of strength and weakness of current models 🤯. Check out the thread below!.
We introduce LiveCodeBench Pro. Models like o3-high, o4-mini, and Gemini 2.5 Pro score 0% on hard competitive programming problems.
0
3
12
RT @arankomatsuzaki: LiveCodeBench Pro: How Do Olympiad Medalists Judge LLMs in Competitive Programming?. - A benchmark composed of problem….
0
26
0
Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.
4
6
37
RT @zzZixuanWang: LLMs can solve complex tasks that require combining multiple reasoning steps. But when are such capabilities learnable vi….
0
37
0
Find us at poster 602 tomorrow morning (10:00-12:30)!.
New unlearning work at #ICLR2025! We give guarantees for unlearning a simple class of language models (topic models), and we further show it's easier to unlearn pretraining data during fine-tuning, without even modifying the base model. Paper: 🧵:.
0
0
4
0
0
1
To summarize: provable unlearning in simple language modeling scenarios is achievable. Our framework paves the way for future theoretical guarantees in more complex, realistic language model settings, beyond topic models. For more details, check out our paper or find us in 🇸🇬!.
1
0
0
An even cooler result: if we only care about downstream utility, we can provably unlearn even more pretraining data while preserving utility! We formalize the intuition that greater task difficulty -> more feature learning -> harder to unlearn.
1
0
0
Our result: an algorithm that outputs an unlearned model that 1) satisfies indistinguishability wrt the retrained* topic model and 2) preserves utility even upon adversarial deletion of training data. *here, we use the learning algorithm from
arxiv.org
Topic models provide a useful method for dimensionality reduction and exploratory data analysis in large text corpora. Most approaches to topic model inference have been based on a maximum...
1
0
0
A prelim on topic models: the topic matrix A represents, for each column (topic), the distribution of words for that topic. Given training documents (a simple bag-of-words representation), goal is to learn the topic matrix A; each document is sampled from some distribution D.
1
0
0
Why unlearning? Well, privacy matters; sensitive data (e.g. phone numbers, personal details) must be removable from trained models without significant performance loss, upon request by a user who previously shared data. Moreover, retraining from scratch is very undesirable.
1
0
0
In practice, we've mastered building powerful LLMs. Yet theoretical guarantees, especially around privacy and unlearning sensitive information, remain elusive. Can we bridge this gap?. Our work: use topic models as a controlled setting to rigorously explore data unlearning.
1
0
0
New unlearning work at #ICLR2025! We give guarantees for unlearning a simple class of language models (topic models), and we further show it's easier to unlearn pretraining data during fine-tuning, without even modifying the base model. Paper: 🧵:.
arxiv.org
Machine unlearning algorithms are increasingly important as legal concerns arise around the provenance of training data, but verifying the success of unlearning is often difficult. Provable...
2
16
67
New insights on understanding reward model selection! Better RM accuracy != better RM for RLHF training; reward variance plays an important role as well.
The success of RLHF depends heavily on the quality of the reward model (RM), but how should we measure this quality?. 📰 We study what makes a good RM from an optimization perspective. Among other results, we formalize why more accurate RMs are not necessarily better teachers!.🧵
0
2
15
RT @parksimon0808: Does all LLM reasoning transfer to VLM? In context of Simple-to-Hard generalization we show: NO! We also give ways to re….
0
18
0
RT @AmartyaSanyal: Open Postdoctoral position in Privacy (and unlearning) and Robustness in Machine Learning in University of Copenhagen to….
0
41
0
Come to us at poster session 406 today from 1:30pm to 3:00pm to chat more!.
Why are transformers more powerful than fully-connected networks (FCNs) on sequential data (e.g. natural language)?. Excited to introduce our #ICML2024 paper: Joint w/ @stanleyrwei, @djhsu, @jasondeanlee (1/n)
0
0
4
At Vienna for ICML the next few days - will be presenting with @zzZixuanWang tomorrow on our most recent work on transformers. Feel free to stop by Hall C 4-9 tomorrow afternoon to check out our work!.
Transformers Provably Learn Sparse Token Selection While Fully-Connected Nets Cannot
6
0
12
RT @StatMLPapers: Transformers Provably Learn Sparse Token Selection While Fully-Connected Nets Cannot
0
3
0