←
Daisuke OBA121 posts
Daisuke OBA avatar

Daisuke OBA

@dai0NLP

Postdoc at @sciencetokyo_en | NLP/ML, Generative models, diffusion LMs, LLM test-time scaling | Ph.D., UTokyo

◷ Joined October 2022▧ 14 Photos and videos
Subham Sahoo avatar
⟳ Daisuke OBA retweeted
Subham Sahoo@ssahoo_·Nov 11

Giving a talk on Diffusion augmented LLMs (Uno) @askalphaxiv Some new results include: > A newer version of Uno which beats DFlash2 > Scaling to 400B scale > Speeding up RL post-training 📅 Friday Nov 13 ⏲️ 10-11am PT Zoom: luma.com/fli4wjlr

Making LLM inference up to 3x faster using discrete diffusion without sacrificing accuracy? Come join us for this AI4Science talk: Unlocking Lossless Speedups in LLMs via Discrete Diffusion. In this session, the author Subham Sahoo will introduce Uno, a new approach from IFM that combines autoregressive and diffusion modeling to accelerate LLM inference. While parallel generation methods often sacrifice accuracy for speed, Uno uses lightweight diffusion adapters to generate multiple tokens in parallel, then relies on the original model to verify them, preserving its output distribution without requiring a separate draft model. So not only is it 3x faster, it also outperforms top speculative decoding methods, with speedups even works at large batch sizes. If you're working or interested in LLM inferencing, this is the talk to check out. :microscope: AI4Science on alphaXiv :calendar_spiral: Friday November 13th 2026 · 10 AM PT :microphone2: Featuring Subham Sahoo :speech_balloon: Casual Talk + Open Discussion Check out the paper here: alphaxiv.org/abs/2609.04010 Sign up for the talk here: luma.com/fli4wjlr

Post attachment
生研マン avatar
⟳ Daisuke OBA retweeted
生研マン@shonosuke_i·Nov 10

Langaku、日本e-learning大賞でまさかのダブル受賞🎉 めでたい!

/ 🏆 2026年度 日本e-Learning大賞 受賞!🏆 \ なんと #Langaku が『大賞』を受賞しました🎉 さらに! 『日本電子出版協会会長賞』も受賞! 嬉しいダブル受賞です✨ いつも温かい応援ありがとうございます! これからも楽しい学習体験をお届けします📚 #eラーニング大賞 #eラーニングアワード

Post attachment
Naoaki Okazaki avatar
⟳ Daisuke OBA retweeted
Naoaki Okazaki@chokkanorg·Nov 10

Rakuten Technology Excellence Awards 2026を岡崎研、横田研、産総研が受賞することになりました! 日本語LLM「Swallow」など、私たちの取り組みを評価していただき大変光栄です。一緒に取り組んできた皆さんに感謝します。引き続き、日本のオープンLLMの研究開発を盛り上げていきたいと思います!

楽天テクノロジー・エクセレンスアワード2026の受賞者が決定! 🏆 JAI-Trust 様 🏆 Apache Iceberg Community 様 🏆 Science Tokyo (Okazaki Laboratory & Yokota Laboratory) and AIST 様 詳細👇 lnky.jp/Bbq844l 11/14開催の #rtechconf 2026にて授賞式をぜひ現地でご覧ください!

Prof. Danushka Bollegala avatar
⟳ Daisuke OBA retweeted
Prof. Danushka Bollegala@Bollegala·Nov 10

@aaclmeeting is coming to an end but we saved the best for the last! Please drop by our tutorial on Diffusion Language Models from 14:00 at Muhe Hall B. The opportunities are immense for the NLP community to contribute to this growing field. Multilinguality, multimodality, interpretability, diversity, safety evals are all pretty much open topics in DLMs and we promise to explain everything that you need to know to get started. slides and reading list at dllm-aacl.github.io/ #diffusionmodels

Yi (Jodie) Zhou avatar
⟳ Daisuke OBA retweeted
Yi (Jodie) Zhou@jodieyzhou·Nov 7

🤔 Do LLMs struggle with word senses, or just make overly fine-grained distinctions? I will present our paper “WiC is Not WSD” at #AACL2026! 📅 Nov 8 | Semantics and Linguistic Structure 📍 Room 5 📄 arxiv.org/abs/2609.20593 Hope to see you there!

Subham Sahoo avatar
⟳ Daisuke OBA retweeted
Subham Sahoo@ssahoo_·Nov 6

Discrete LLM Tutorial at AACL 2026 🔥

Post attachment

I’ll be giving a tutorial on diffusion language models at AACL-IJCNLP 2026 on Nov. 10 w/ @Bollegala 🔗 dllm-aacl.github.io/ (Slides coming soon!) #aacl2026 @aaclmeeting 💭Grateful for the opportunity to organize my perspective on this fast-moving field and share it with the NLP community. This will be my first tutorial, so it’s a bit of a personal challenge for me as well. Looking forward to the discussions :)

Daisuke OBA avatar
Daisuke OBA@dai0NLP·Nov 6

I’ll be giving a tutorial on diffusion language models at AACL-IJCNLP 2026 on Nov. 10 w/ @Bollegala 🔗 dllm-aacl.github.io/ (Slides coming soon!) #aacl2026 @aaclmeeting 💭Grateful for the opportunity to organize my perspective on this fast-moving field and share it with the NLP community. This will be my first tutorial, so it’s a bit of a personal challenge for me as well. Looking forward to the discussions :)

Prof. Danushka Bollegala avatar
⟳ Daisuke OBA retweeted
Prof. Danushka Bollegala@Bollegala·Nov 5

✈️ Off to Zhuhai for #AACL2026! 🚨 We are hiring! I’m looking for NLP talent at all levels: 🏛️ @LivUniCS (PhD students to Full Prof in AI) 📦 Amazon (Applied Scientists) My DMs are open—let's catch up on roles or research! Catch our work this week 👇 🗓️ Nov 7 | 10:30am (Interpretability) Oral: Decoupling Internal Representational Changes & Causal Importance in Fine-Tuned LLMs 🔗 aclanthology.org/2026.aacl-l… w/ @prochetasen @li_lingfang Subham Das 🗓️ Nov 8 | 10:30am (Semantics) Oral: WiC is not WSD: A Study on LLMs & Lexical Ambiguity Resolution 🔗 aclanthology.org/2026.aacl-l… w/ @jodieyzhou Kiamehr Rezaee @tpilehvar @CamachoCollados 🗓️ Nov 10 | 2:00pm Tutorial: Diffusion Language Models 🔗 dllm-aacl.github.io/ w/ @dai0NLP See you there! @aaclmeeting #aacl2026 #LiverpoolNLP @AmazonScience

Yuki Otsuka avatar
⟳ Daisuke OBA retweeted
Yuki Otsuka@yukiot22·Oct 24

Drifting Models are one-step generators. Can we give them an iterative refinement mechanism for text generation? We propose ReLaDrift: a Drifting Model drafts from noise, then recurrently refines its own draft in latent space. It works on Sudoku and GSM8K! w/ @dai0NLP @chokkanorg

1:08
Stefano Ermon avatar
⟳ Daisuke OBA retweeted
Stefano Ermon@StefanoErmon·Oct 19

1/4 Today, we’re releasing Mercury 3. Mercury 3 is by far our most powerful model and the most intelligent diffusion LLM ever released. It outperforms GPT-6 Luna and Haiku 5.5 on quality while running at 1000+ tok/sec.

0:30
Prof. Danushka Bollegala avatar
⟳ Daisuke OBA retweeted
Prof. Danushka Bollegala@Bollegala·Sep 11

Thrilled to announce that we have two papers and a tutorial accepted to AACL-IJCNLP. 🎉 1. Decoupling Internal Representational Changes and Causal Importance in Fine-Tuned Large Language Models (with Lingfang Li, Procheta Sen, and Shubham Das) We investigated how fine-tuning fundamentally alters internal representation (like attention patterns and layer-wise activations) inside LLMs and discovered that components most causally important to a task's performance are actually concentrated within specific layers. Interestingly, the distribution of these causally important components is largely uncorrelated with the layers undergoing the most substantial representational changes during fine-tuning. If you are interested in how fine-tuning affects a model internally LLMs please get in touch. P 2, WiC is Not WSD: A Study on LLMs and Lexical Ambiguity Resolution (with @jodieyzhou , Kiamehr Rezaee, Mohammad Taher Pilehvar, and @CamachoCollados ) We explored why Word-in-Context (WiC) tasks remain so challenging for LLMs compared to traditional Word Sense Disambiguation (WSD). The difficulty isn't just about comparing two contextual uses of a word, but rather the lack of an explicit sense inventory that specifies the required level of semantic granularity. Providing candidate senses—similar to what is done in traditional WSD—improves WiC performance across all settings by helping models make more consistent judgments. Our human evaluation showed that many apparent LLM errors aren't simple failures of lexical understanding. Instead, LLMs tend to "overthink" and apply overly fine-grained sense distinctions compared to human annotators. 3. I will be giving a tutorial on Diffusion Language Models with @dai0NLP at AACL as well Preprints for both papers will be available soon. Congratulations to my collaborators/students #aacl26 #NLProc #LiverpoolNLP