Berk Ustun

Denied a loan by an ML model? You should be able to change something to get approved! In a new paper w @AlexanderSpangh & @yxxxliu, we call this concept "recourse" & we develop tools to measure it for linear classifiers. PDF bit.ly/2xh0idm CODE bit.ly/2xiEvls

Murderers

Neurips 2025 was such a blast! We snuck a grand piano into the CreativeAI Track to demo Aria, our pretrained chat-style music model:

A 'pragmatic intepretability' turn sounds a lot like our argument/framework for evaluating explanation methods--Time to replace task-agnostic fortunetelling w/concrete decision problem specs + theoretic & empirical evidence of expected performance boost arxiv.org/abs/2506.22740

GPT-5 on Sudoku-Bench 🧩 Since releasing Sudoku-Bench in May 2025, when no LLM could solve a classic 9x9 puzzle, we've been evaluating the latest generation of models. GPT-5 now leads our leaderboard with 33% puzzles solved--approximately 2x the previous leader--and is the first LLM we've tested to solve a 9x9 Sudoku variant. However, with 67% of the much harder puzzles remaining unsolved, Sudoku-Bench continues to present significant challenges for AI reasoning. Modern Sudoku variants require models to first understand novel rulesets through meta-reasoning, then maintain global consistency across long reasoning chains. Our experiments with GRPO fine-tuning on Qwen2.5-7b and "Thought Cloning" (training on expert human reasoning from Cracking the Cryptic) show that current approaches still struggle with the spatial reasoning and creative "break-in" points that human solvers use naturally. We believe new approaches are required to solve our benchmark. These results highlight persistent gaps between computational problem-solving and human-like reasoning, particularly in tasks requiring integrated mathematical logic, spatial awareness, and creative insight. Read more about our update here: 🔗 Blogpost → pub.sakana.ai/sudoku-gpt5/

✨ Very overdue update: I'll be starting as an Assistant Professor in CS at University of Minnesota, Twin Cities, Fall 2026. I will be recruiting PhD students!! Please help me spread the word! [Thread] 1/n

About a week away from the deadline to submit to the ✨ Workshop on Algorithmic Collective Action (ACA) ✨ acaworkshop.github.io at NeurIPS 2025!

I often wonder whether the prospective grad students who contact me understand what they are signing up for. I hope this does the trick

PhD in Computer Science, University of California San Diego 🎓 My research focused on uncertainty and safety in AI systems, including 🤷‍♀️letting models say "I don't know" under uncertainty 🔎understanding and reducing hallucinations 🔁 methods for answering "how much will providing data X improve performance on Y?" at inference time Many thanks to my advisor @berkustun, to my incredible research collaborators, and to my wonderful friends, husband and family. Getting a PhD while becoming a first-time parent is definitely a recipe for growth!

Our new ICML 2025 oral paper proposes a new unified theory of both Double Descent and Grokking, revealing that both of these deep learning phenomena can be understood as being caused by prime numbers in the network parameters 🤯🤯 🧵[1/8]

Explainable AI has long frustrated me by lacking a clear theory of what explanations should do. Improve use of a model for what? How? Given a task what's max effect explanation can have? It's complicated bc most methods are functions of features & prediction but not true state 1/

Explanations don't help us detect algorithmic discrimination. Even when users are trained. Even when we control their beliefs. Even under ideal conditions... 👇

We’ll be presenting @FAccTConference on 06.24 at 10:45 AM during the Evaluating Explainable AI session! Come chat with us. We would love to discuss implications for AI policy, better auditing methods, and next steps for algorithmic fairness research. #AIFairness #xAI

Right to explanation laws assume explanations help people detect algorithmic discrimination. But is there any evidence for that? In our latest work w/ David Danks @berkustun, we show explanations fail to help people, even under optimal conditions. PDF shorturl.at/yaRua

💡LLMs don’t have to be black boxes. We introduce CB-LLMs -- the first LLMs with built-in interpretability for transparent, controllable, and safer AI. 🚀Our #ICLR2025 paper: lilywenglab.github.io/CB-LLMs/ #TrustworthyAI #ExplainableAI #AI #MachineLearning #NLP #LLM #AIResearch

When RAG systems hallucinate, is the LLM misusing available information or is the retrieved context insufficient? In our #ICLR2025 paper, we introduce "sufficient context" to disentangle these failure modes. Work w J. Zhang, C.S. Ferng, @DaChengJuan1, @ankurtaly @CyrusRashtchian

Denied a loan, an interview, or an insurance claim by machine learning models? You may be entitled to a list of reasons. In our latest w @anniewernerfelt @berkustun @kdphd, we show how existing explanation frameworks can fail and present an alternative tailored for recourse 🧵

Why is it so hard to show that people can be better decision-makers than statistical models? Some ways that common intuitions about the superiority of human judgment contradict statistical reality, and a few that don't. statmodeling.stat.columbia.edu/2025/04/18/dum…

Many ML models predict labels that don’t reflect what we care about e.g.: – Diagnoses from unreliable tests – Outcomes from noisy electronic health records In our #ICLR2025 paper, we study how this subjects individuals to a lottery of mistakes Paper: bit.ly/3Y673uZ 🧵👇

🚨 Excited to announce a new paper accepted at ICLR2025 in Singapore! “Learning Under Temporal Label Noise” We tackle a new challenge in time series ML: label noise that changes over time 🧵👇 arxiv.org/abs/2402.04398

🔥🔥Our small team in Seattle Google DeepMind is hiring! 🔥🔥If you are willing to move to/already in Seattle, has done significant work on human-machine communication / interpretability (from ML side) with a relevant PhD and great publication record, Join us. Apply here 👉👉 boards.greenhouse.io/deepmind/jobs/…