YJ Choe (@_Mr_YJ) profile picture

Asst Prof @INSEAD Decision Sciences. PhD from @CMU_StatDS & @mldcmu; former postdoc @DSI_UChicago. Statistics, forecasting, ML & LLMs

Singapore
Joined January 2016
Photo by YJ Choe (@_Mr_YJ)Photo by YJ Choe (@_Mr_YJ)Photo by YJ Choe (@_Mr_YJ)
📢 The e-value workshop at #NeurIPS2026 (Paris) is accepting submissions! e-values-workshop.github.io/ We especially welcome early-stage work connecting e-value theory to modern ML topics, such as conformal prediction, online learning, and AI auditing. 🗓️ Deadline: 29 August AoE
1
1
2
291
"E-values: From Statistics to ML" 🎤 Invited speakers: Michael Jordan, Rianne de Heide, Eugenio Clerico, Emille Kaufmann, and Nick Koning The workshop will take place on either December 12 or 13 in the Paris satellite event of NeurIPS 2026.
1
3
329
The poster presentation continues today! @KihoPark_ @toddknife @victorveitch
I'll be presenting our paper on information geometry at ICML! 📌 Thu (July 9), 5 PM, Hall A, Poster #3301 📌 Fri (July 10), Hall C (Mech Interp Workshop) Come say hi if you're at ICML! Happy to discuss our work or grab a coffee and chat :)
1
245
This paper dramatically changed my understanding of how to think about the internal representations of models. Check it out :)
Interpreting and controlling internal representations should be based on how the model actually uses them! Turns out: information geometry makes this precise. We show how, and use it to derive a (provably & empirically) robust strategy for steering. arxiv.org/abs/2602.15293
1
15
287
38,289
Quite excited about our new preprint on representation geometry in LLMs/VLMs! Information geometry provides a powerful formalism for making sense of the high-dim embedding spaces in these softmax-based models. Most of the credit goes to @KihoPark_ @toddknife @victorveitch.
Interpreting and controlling internal representations should be based on how the model actually uses them! Turns out: information geometry makes this precise. We show how, and use it to derive a (provably & empirically) robust strategy for steering. arxiv.org/abs/2602.15293
7
453
Semantics in language is naturally hierarchical, but attempts to interpret LLMs often ignore this. Turns out: baking semantic hierarchy into sparse autoencoders can give big jumps in interpretability and efficiency. Thread + bonus musings on the value of SAEs:
11
60
5
298
27,401
This is happening today! Come say hi at the poster from 10:00-12:30 😄 If not, the talk is in the late afternoon at 16:30 #ICLR2025
Tomorrow (Fri) at #ICLR2025, our paper on the geometry of categorical and hierarchical concepts in LLMs will be presented in oral (4:30 PM, Garnet 216–218) and poster (10 AM, Hall 3 + Hall 2B #525) sessions. Sadly, I can’t be there in person.. but @_Mr_YJ will be presenting!
1
8
494
I am in Nice this week attending #ICSDS2024! Slides from my talk on e-values/e-processes & filtrations are now available on my webpage: yjchoe.github.io/slides/Mind… Happy to chat with anyone at the conference interested in e-values/processes, forecasts, or LLM transparency :)
6
361
My blog post on e-values, game-theoretic stats, & recent work on "cooperative skepticism" is now live on the @DSI_UChicago webpage! (This is a short, informal intro; if you'd like to learn more, you can find more details about the paper here: yjchoe.github.io/research/ch…)
What’s game-theoretic about statistics? DSI Postdoctoral Fellow Yo Joong “YJ” Choe introduces recent work developing a general approach to combining game-theoretic evidence in the challenging case of asymmetric information in a new DSI Insights Blog! datascience.uchicago.edu/ins…
1
2
12
1,216
I am at #INFORMS2024 to present my work on evaluating black-box abstaining classifiers! Time: 1:45pm-2pm, 10/22 (Tue) Place: Summit - Terrace Suite 1 Session: “TC57 - Causal Inference” Happy to connect if you are at the conference 😄 arxiv.org/abs/2305.10564
Tomorrow (Tue) 5:15pm at #NeurIPS2023, I'm presenting my poster on how to evaluate & compare black-box abstaining classifiers. Takeaway: when classifiers hide their predictions selectively, the problem should be viewed as a causal inference problem. Learn more at poster #1618!
1
4
1,046
🧵 RATE: Score Reward Models with Imperfect Rewrites of Rewrites 1/ How do you measure whether a reward model incentivizes helpfulness without accidentally measuring length, complexity, etc? Rewrites of rewrites give good counterfactuals, without needing to list all confounders!
1
10
1
15
2,077
Exciting news about our LLM work from last week's workshop at ICML! Kudos to @KihoPark_ for leading this one!
Very happy that this won best paper at the MI workshop :) Great moment to follow @KihoPark_ @_Mr_YJ @yibophd to keep track of all the great stuff they're doing!
1
4
435
For the Stats crowd: unlike P-values, E-values (en.wikipedia.org/wiki/E-valu…) can be combined easily under arbitrary dependence. But can we say the same for their sequential analog, e-processes? Sadly the answer is *no*, motivating our recent work: arxiv.org/abs/2402.09698 (A thread)
1
4
1
13
4,802
The e-lifting result implies that, if we have a powerful e-process in a sub-filtration, then we also have a powerful e-process in the data filtration. We also show how an adjuster may be necessary to lift arbitrary e-processes: any e-lifter of its form *must* be an adjuster.
1
289
For more details, see our - Preprint: arxiv.org/abs/2402.09698 - Slides: yjchoe.github.io/slides/Choe… This is joint work with the wonderful Aaditya Ramdas from @CMU_StatDS. I'm presenting this work at the upcoming #JSM2024 in August and #ICSDS2024 in December. Hope to see you there!
2
2
423
For the LLM crowd: Our new paper studies how LLMs represent categorical concepts (as simplices) and hierarchical relations (as orthogonality)! Joint work with @KihoPark_ @yibophd @victorveitch, @DSI_UChicago. arXiv: arxiv.org/abs/2406.01506
Fundamentally, high-level concepts group into categorical variables---mammal, reptile, fish, bird---with a semantic hierarchy---poodle is a dog is a mammal is an animal. How do LLMs internally represent this structure? arxiv.org/abs/2406.01506
4
15
1,489
LLM best-of-n sampling works great in practice---but why? Turns out: it's the best possible policy for maximizing win rate over the base model! Then: we use this to get a truly sweet alignment scheme: easy tweaks, huge gains w @ybnbxb @ggarbacea arxiv.org/abs/2406.00832
5
20
1
83
16,250