@levelfour_

blue sky → https://xitter.cf/t.co/Uj3oDOwSRj

hermite.jp/
Joined March 2010
  • Tweets3.7K
  • Following675
  • Followers1.8K
359Media
ibis2026ではインターナショナルデー&女性向け交流会の企画があります 交流しやすいコミュニティを目指します、みなさまどうぞご参加ください ibisml.org/ibis2026/registra… ibisml.org/ibis2026/women-in…
5
4
1,147
凸解析 × モデル検査の論文を書きました。まさか形式検証の研究（の端くれ）をするとは思いだにしませんでした。検証の分野にこんなにリッチな凸性があることを教えてくれた @KWatanabeCS と、相変わらずスマートな凸解析のアイデアに貢献してくれた @shinsaku_sakaue との共同研究です。
HB@levelfour_Oct 14
Convex analysis meets model checking🧐 Probabilistic model checking can be seen as set inclusion test between a safe region and the achievability set to failure states in MDP. Using convex analysis, we propose a computationally efficient query. arxiv.org/abs/2610.17541
17
1
80
5,274
Convex analysis meets model checking🧐 Probabilistic model checking can be seen as set inclusion test between a safe region and the achievability set to failure states in MDP. Using convex analysis, we propose a computationally efficient query. arxiv.org/abs/2610.17541
1
11
1
60
7,927
New paper with my ISM colleague A. Okazaki. Empirical Bayes, with roots in the 1950s, and the Mixture of Experts familiar from LLMs meet in a single estimation problem. arxiv.org/abs/2610.08321
2
3
1,011
「若手交流企画」の参加登録期間を延長しました！ ・参加登録〆切: 2026/10/9(金) ・参加登録フォーム: docs.google.com/forms/d/e/1F… この機会にぜひ参加をご検討ください！

IBIS2026 若手交流企画 参加申し込み

IBIS2026 若手交流企画 (https://ibisml.org/ibis2026/newinibis/) の参加申し込みページです。 【申し込みに関する注意点】 ポスター発表への参加には別途IBIS2026 ワークショップ (https://ibisml.org/ibis2026/) への参加登録が必須です． お申込みに関してご質問のある場合は [email protected]...

docs.google.com
情報論的学習理論と機械学習：IBISML@ibismlAug 25
IBIS2026の「若手交流企画」ページを公開しました！ 本企画は学生の皆様を対象に、研究紹介などを通じて交流を深めるためのプログラムです。 IBISに初めて参加される方も、さらにネットワークを広げたい方も大歓迎です。皆様のご参加をお待ちしております！ibisml.org/ibis2026/newinibi…
9
10
3,644
We are excited to have following papers at #NeurIPS2026! I'm proud of all of them, but in particular, it's fantastic to have the first collaboration with my first PhD student @XianliangLi910 on momentum analysis of Muon. Come to see his work 👉 xitter.cf/XianliangLi910/status/…
Xianliang Li@XianliangLi910Jun 3
🧠Why does Muon do momentum before orthogonalization? ✨Our key insight: momentum acts as a spectral filter for the matrix-valued gradient, yielding a more reliable update for the orthogonalization step. 📝Paper: arxiv.org/abs/2606.03899 🌐Project: yinleung.com/denoise-ortho

Denoise First, Orthogonalize Later: Understanding Momentum in Muon...

Muon has recently demonstrated strong empirical performance in large language model training, but the theoretical role of momentum in Muon remains unclear. Existing analyses of Muon either remove...

arxiv.org
1
6
72
6,025
I got to know this work recently, which is based on a very cool formulation of layer-wise + learning dynamics of Transformer as an optimal control problem!
しんくん@nobo0409Sep 25
"We are delighted to inform you that your submission, "Training-Induced Escape from Token Clustering in a Mean-Field Formulation of Transformers"(arxiv.org/abs/2605.07772), has been accepted at NeurIPS 2026 Main Track as poster. Congratulations!"

Training-Induced Escape from Token Clustering in a Mean-Field...

Transformers perform inference by iteratively transforming token representations across layers. This layerwise computation has been studied empirically, and recent mean-field theories of...

arxiv.org
1
12
3,412
Very excited to share that both of my submissions have been accepted to #NeurIPS2026!! 🎉🎉 And one of them was selected for an oral presentation! 🥳✨ Huge thanks to Yuichi @oxytan for the wonderful collaboration! 🙏 Looking forward to NeurIPS! 🚀
3
7
113
7,911
Recent distortion theory argues RLHF is exponentially suboptimal to align with heterogeneous user utilities, but why do people remain to rely RLHF? We showed RLHF's exponential deterioration can be mitigated by taking distribution mismatch (B) into account.
Kazusato Oko@kazusokoSep 18
Finally uploaded our ICML paper analyzing RLHF through the lens of social choice theory. When users have heterogeneous preferences, RLHF may fail to maximize average user utility. We revisit prior worst-case results and identify conditions under which RLHF works well.
2
2
27
3,691
Here is the link to arXiv arxiv.org/abs/2609.12651 @kazusoko contributed quite a lot and we have presented it at #ICML2026

Distortion of AI Alignment Revisited: RLHF is a Decent Utilitarian Aligner

While Reinforcement Learning from Human Feedback (RLHF) is the standard paradigm for aligning large language models with human preferences, its effectiveness in pluralistic settings has been...

arxiv.org
445
24 Fields Medal winners have written a letter of objection economist.com/science-and-te…

Top mathematicians are outraged by OpenAI’s methods

A group of Fields Medal winners have written a letter of objection

economist.com
67
210
51
827
574,808
学習ダイナミクスに関する2日の連続講演を終えました。長丁場にお付き合いいただいた皆さまありがとうございました。線形NNの力学系、表現学習、安定性の縁、暗黙的正則化、振動力学系等まとめていますので、参考にしていただければと思います。 speakerdeck.com/hanbao/rikig…

勾配流から乖離した学習ダイナミクス

数理最適化で繋ぐ幾何解析と応用数学 第1回（ https://sites.google.com/g.ecc.u-tokyo.ac.jp/26h01996/%E7%AC%AC1%E5%9B%9E ）の連続講演で用いた資料です。

speakerdeck.com
30
163
7,543
来週 9/9 の OR 学会シンポジウム で「逆最適化と機械学習」というタイトルで講演します！orsj.org/2026f/symposium/ スライドはこちら👉 speakerdeck.com/ssakaue/gyak…
16
1
57
5,307
I'm highly thrilled to host Mohit as a new postdoc in my team! He has a solid series of work on the foundation of fairness in ML. Mohit Sharma さんが研究員としてジョインしました。機械学習の公平性をはじめ、博士時代に良い仕事をされているので、これからの共同研究が楽しみです。
Mohit Sharma@mohitsharmaj29Sep 2
Replying to @mohitsharmaj29
I have now joined as a Postdoc at Tohoku University and the Institute of Statistical Mathematics Tokyo, where I will work with Prof. Han Bao (@levelfour_ ). (2/2)
25
2,871
宣伝: 9/9〜9/11 に名古屋大学で開催される【数理最適化で繋ぐ幾何解析と応用数学】という会合（sites.google.com/g.ecc.u-tok…）で、学習ダイナミクスに関する3時間の講演を行います。このスライド（speakerdeck.com/hanbao/li-xu…）の周辺の話を数理的に詳しく喋ります。まだ参加登録中ですのでよろしければ。

数理最適化で繋ぐ幾何解析と応用数学 - 第1回

参加登録：https://forms.gle/SsoBEHTsRiFER3VV9 （懇談会の申込締切：2026年8月31日(月)） 場所：名古屋大学　多元数理科学棟 509講義室（アクセス） 日時：2026年9月9(水)〜9月11日(金) ポスターのDLはこちらからもできます。

sites.google.com
26
80
6,977
統数研プロジェクト紹介「第46回：最適輸送理論でAIを進化させる」を掲載 ism.ac.jp/ism_info_j/labo/pr…
27
100
12,465
IBIS2026のCFPを公開しました。今年は新たな試みとして、英語でのポスター発表枠を用意しています。 ・日時：2026年10月26日～29日 ・会場：ソニックシティ（埼玉） 追加情報は決定次第、更新していきます。ibisml.org/ibis2026/
24
3
36
9,289
🎓 Confirmed speakers for MLSS Okinawa 2027! 🇯🇵 📍 Onna Village, Okinawa, Japan 🗓️ March 1–12, 2027 Pierre Alquier (ESSEC) — PAC-Bayes Theory Randall Balestriero (Brown) — World Models (JEPA) Shai Ben-David (Waterloo / Vector Institute) — Learning Theory Danqi Chen (Princeton) — Language Models / NLP Marco Cuturi (Apple / ENSAE) — Optimal Transport Arnaud Doucet (Google DeepMind) — Diffusion Models Arthur Gretton (UCL) — Kernel Methods Bo Han (Hong Kong Baptist University) — Trustworthy ML Junya Honda (Kyoto University) — Bandits Gintare Karolina (Google DeepMind) — Data Attribution Peter Richtarik (KAUST) — Distributed Optimization Taiji Suzuki (U. Tokyo / RIKEN AIP) — Statistical Learning Theory Frank Wang (NTU / NVIDIA) — Computer Vision Sherry Yang (NYU / Google DeepMind) — Representation Learning Han Zhao (UIUC) — Robust Machine Learning More speakers to be announced soon! 📝 Applications are open now — no application fee. Apply: openreview.net/group?id=OIST… Info: oist.github.io/mlss2027okina… #MachineLearning #MLSS #Okinawa #OIST

OIST 2027 MLSS Okinawa

Welcome to the OpenReview homepage for OIST 2027 MLSS Okinawa

openreview.net
7
2
54
11,441
Non-asymptotic implicit bias of logistic regression at early-stage gradient descent dynamics ift.tt/b0HjfTk

Non-asymptotic implicit bias of logistic regression at early-stage...

Gradient descent has been of particular interest in modern machine learning beyond sole focus on optimization. Implicit bias emerging from optimization, though not being encoded by the learning...

arxiv.org
8
22
2,140
My latest work: GD's max-margin bias is well-known, but its conv rate is very slow. Still we can observe decent convergence in practice (see fig). I have been thinking it for a while and figured out early weak alignment is possible (with a clean proof!) arxiv.org/abs/2608.04382
5
48
4,498
Load more