Maya Okawa

Research Scientist at CBS-NTT Program in Physics of Intelligence, Harvard University.

[Bo Zhao @BoZhao__]

How do LLMs organize emotion concepts?

Excited to share our #ICML2026 paper: we uncover emergent “emotion trees” in LLMs, inspired by emotion wheels from psychology.

📄Emergence of Hierarchical Emotion Organization in Large Language Models
🔗https://t.co/6RxrH9pMiN

[Yongyi Yang @YongyiYang7]

What drives in-context learning in LLMs?

New paper: Provable Low-Frequency Bias of In-Context Learning of Representations.

We show LLMs have a low-frequency bias when learning representations in context, offering a theoretical answer to several previously open questions. 🧵👇

[Ekdeep Singh Lubana @EkdeepL]

Paper link: https://t.co/Pg1ujFhkr1

Shoutout especially to our monster undergrad @kento_nishi who led this project, and all collaborators!
@MayaOkawa, @RahulRam3sh, Mikail Khona, and @Hidenori8Tanaka

[Kento Nishi｜🐔 @kento_nishi]

🚨 ICML 2025 Paper! 🚨

Excited to announce "Representation Shattering in Transformers: A Synthetic Study with Knowledge Editing."

🔗 https://t.co/If0OHOWSJn

We uncover a new phenomenon, Representation Shattering, to explain why KE edits negatively affect LLMs' reasoning.

🧵👇

[Ekdeep Singh Lubana @EkdeepL]

🚨 New paper alert!

Linear representation hypothesis (LRH) argues concepts are encoded as **sparse sum of orthogonal directions**, motivating interpretability tools like SAEs. But what if some concepts don’t fit that mold? Would SAEs capture them? 🤔

1/11

[Ekdeep Singh Lubana @EkdeepL]

New paper---freshly accepted to ICML!

Detailed thread coming soon, but pretty excited about this project. We use synthetic knowledge graphs to study why knowledge editing protocols can screw up model capabilities, finding what we call a "representation shattering" effect!

[Hidenori Tanaka @Hidenori8Tanaka]

Q. Why does editing knowledge sometimes reduce overall performance?

A. Representation shatters!

Great insights by a super start undergrad @kento_nishi, supervised expertly by @EkdeepL, with @MayaOkawa, @RahulRam3sh, and Mikail Khona.

[Ekdeep Singh Lubana @EkdeepL]

New paper–Accepted at #ICLR2025 and also my last PhD paper! 🧑‍🎓🧵👇

We propose a novel model of how emergent learning curves show up in neural nets’ training by making a connection to the theory of graph percolation!

[Maya Okawa @MayaOkawa]

🎉Excited to announce that our new paper on learning dynamics in concept graphs has been accepted as a #NeurIPS2024 spotlight!
https://t.co/vXV2jFg6YR

[Ekdeep Singh Lubana @EkdeepL]

Paper alert—accepted as a NeurIPS *Spotlight*!🧵👇

We build on our past work relating emergence to task compositionality and analyze the *learning dynamics* of such tasks: we find there exist latent interventions that can elicit them much before input prompting works! 🤯

[Hidenori Tanaka @Hidenori8Tanaka]

Excited to share our #NeurIPS2024 spotlight!
https://t.co/XEAnrWo0XP

Key finding: Modern multi-modal models may be hiding far greater capabilities than current protocols can reveal.

Challenge: How do we unlock these "hidden capabilities" through novel extraction protocols?

[Ekdeep Singh Lubana @EkdeepL]

Paper alert—accepted as a NeurIPS *Spotlight*!🧵👇

We build on our past work relating emergence to task compositionality and analyze the *learning dynamics* of such tasks: we find there exist latent interventions that can elicit them much before input prompting works! 🤯

[DLHacks @DL_Hacks]

【松尾研Diffusion勉強会資料公開】

拡散モデルの構成的汎化能力を検証するための概念グラフというフレームワークを提案．どういった条件下で構成的汎化に失敗するかを検証

https://t.co/R0hBzD7x5y

[Maya Okawa @MayaOkawa]

@YouTube 論文はこちらです
https://t.co/gkkJwGnhSI

[Maya Okawa @MayaOkawa]

#NeurIPS2023 論文"Compositional Abilities Emerge Multiplicatively: Exploring Diffusion Models on a Synthetic Task"のビデオがYouTube で公開されました。ぜひ見てください！
Diffusion Model - Maya Okawa https://t.co/bvhBYbI5Pc via @YouTube

[Maya Okawa @MayaOkawa]

Read the full paper here: https://t.co/gkkJwGnhSI

[Maya Okawa @MayaOkawa]

Our NeurIPS 2023 paper, "Compositional Abilities Emerge Multiplicatively: Exploring Diffusion Models on a Synthetic Task," now has a video on YouTube!
It's in Japanese, but English subtitles are available. Check it out: https://t.co/MSm5ZBpFDa

[Maya Okawa @MayaOkawa]

Read the full paper here: https://t.co/gkkJwGnhSI

[Maya Okawa @MayaOkawa]

Our NeurIPS 2023 paper, "Compositional Abilities Emerge Multiplicatively: Exploring Diffusion Models on a Synthetic Task," now has a video on YouTube!
It's in Japanese, but English subtitles are available. Check it out: https://t.co/MSm5ZBpFDa

[Hidenori Tanaka @Hidenori8Tanaka]

More excited than ever to announce $1.7M:
"CBS-NTT Program in Physics of Intelligence at Harvard"! 🧠

With new technology comes new science. The time is ripe to build a better future with "Physics of Intelligence for Trustworthy and Green AI"!

🧵👇https://t.co/60arfZrQAs

[Hidenori Tanaka @Hidenori8Tanaka]

ハーバード大学に、"Physics of Intelligence" (知性の物理学) プログラムを新設しました！

「知性とは何か？」という問いに、脳だけでなく生成AI等も実験系として用いることで理論構築に挑みます。

得られた知見を、信頼性が高く環境に優しいAIの開発に応用します。

https://t.co/Z2E02CtUPk

[Ekdeep Singh Lubana @EkdeepL]

Quick shoutout to my awesome collaborators on this! :)
@MayaOkawa, Robert Dick, and @Hidenori8Tanaka