Atsuyuki Miyai @UTokyo

190 posts
Opens profile photo
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
2nd-year PhD student at The University of Tokyo 🇯🇵, Yamasaki Lab and Aizawa Lab | Computer Vision | Trustworthy AI | Account for Research Purposes
The University of Tokyo, Japanatsumiyai.github.io
Joined July 2023
460 Following
569 Followers

Atsuyuki Miyai @UTokyo’s posts

Pinned
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
✨God has shown the path to Auto Research! If you're starting work in this area, check out the latest AI Scientist paper (TMLR2026)! Our paper explores Auto Research from a similar perspective and helps clarify promising research directions and future work! Jr. AI Scientist
Image
Image
Quote
Andrej Karpathy
@karpathy
Image
I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then: - the human iterates on the
25K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
🧙‍♂️ Imagine web agents that don’t just browse but handle your tedious digital chores! 📣 Our team developed WebChoreArena - 532 human-curated tasks, crafted over 300+ hours - Tests agents on massive information memorization, mathematical reasoning, and long-term memory -
Image
18K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
東大の相澤研でJMMMU （LMMの日本語用のベンチマーク）を作りました! 相澤研は今年度でなくなってしまうのですが、研究室がなくなっても日本の画像のコミュニティと言語のコミュニティの両方で長く使われればいいなという願いも込めて作りました! 是非研究開発にご利用ください!
Quote
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
⭐️ Ready for the next stage of multi-lingual LMM🌏? 📣 Happy to share our JMMMU🇯🇵, a Japanese MMMU benchmark! For many users, it’s important to accelerate non-English research. JMMMU will accelerate research in Japanese and multi-lingual LMMs! HP: mmmu-japanese-benchmark.github.io/JMMMU/
Image
26K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
🎉Our survey on how OOD detection & related tasks have evolved in the VLM and Large VLM era is accepted to #TMLR! The field is finally coming together, and OOD detection & anomaly detection are now at the center in the VLM era. In the LVLM era, UPD (Unsolvable Problem
Image
13K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
軽く宣伝です! 明日の15:30から東大二号館で日本のCVを引っ張ってきた相澤清晴教授の最終講義が行われます。 事前予約されていない場合でも、当日参加分の枠が先着順でいくつかあるので、参加可能かと思います。 お時間ある方は是非！ 時間: 15:00~ 開場、15:30~講義 場所: 工学部二号館213教室
Image
13K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
#NeurIPS2023, “LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning”の論文を更新しました。CoOpに背景などの情報を分布外データとして学習するロスを加えるだけで、ImageNetベンチマークにおいて1ショット (15分の訓練)でも既存のCLIPベースの手法を超える精度を達成しました！！
Image
7.7K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
🎉April Fool is the birthday of Unsolvable Problem Detection! UPD examines the VLM’s ability to withhold answers when faced with unsolvable problems. Please enjoy VLMs with unsolvable problems today! paper page: arxiv.org/abs/2403.20331 code: github.com/AtsuMiyai/UPD
Quote
AK
@_akhaliq
Unsolvable Problem Detection Evaluating Trustworthiness of Vision Language Models This paper introduces a novel and significant challenge for Vision Language Models (VLMs), termed Unsolvable Problem Detection (UPD). UPD examines the VLM's ability to withhold answers when
Image
35K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
🎊 Our JMMMU has been accepted by #NAACL2025 main conference! Kudos to the hard work of our great coauthors, reviewers, ACs, and many volunteers 🙌
Quote
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
⭐️ Ready for the next stage of multi-lingual LMM🌏? 📣 Happy to share our JMMMU🇯🇵, a Japanese MMMU benchmark! For many users, it’s important to accelerate non-English research. JMMMU will accelerate research in Japanese and multi-lingual LMMs! HP: mmmu-japanese-benchmark.github.io/JMMMU/
Image
5.3K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
Happy to share that our work "LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning" has been accepted to #NeurIPS2023! 🎉 I would like to express my deepest appreciation to my co-authors and reviewers. arXiv:
arxiv.org
LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning
We present a novel vision-language prompt learning approach for few-shot out-of-distribution (OOD) detection. Few-shot OOD detection aims to detect OOD images from classes that are unseen during...
14K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
(Just my personal opinion,) I wish CV conferences could consider introducing a “Data-centric/Application Track,” similar to ICML or NeurIPS. In my experience, CV conferences tend to equate novelty with proposing a specific approach to a specific problem, more so than in NLP or
4.2K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
✈️I’ll attend #NeurIPS2024 and have an oral presentation@EvalEval Workshop🤗 🕐 15th, 11:30 AM - 12:30 PM Feel free to drop by if you can☕️ Looking forward to exploring the intersection of diverse fields and catching up with the latest community trends! 🚀🚀
Quote
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
⭐️ Ready for the next stage of multi-lingual LMM🌏? 📣 Happy to share our JMMMU🇯🇵, a Japanese MMMU benchmark! For many users, it’s important to accelerate non-English research. JMMMU will accelerate research in Japanese and multi-lingual LMMs! HP: mmmu-japanese-benchmark.github.io/JMMMU/
Image
3.1K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
昨年の卒論生の川上君の仕事がMIRUオーラルに選ばれました! Machine Unlearningは著作権保護のための分野でアカデミアの役割の一つだと思ってます ただ、画像言語の領域だと分野が確立されていない事に課題を感じ、評価方法を模索しました 川上君や
@kazukiega
が主に頑張ってくれました! 是非!
Image
7.2K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
Unfortunately, I missed NAACL2025 (yes, my research budget limits…🤣), but
@ku21fan
will be there in person to present our JMMMU poster! This is the very first step in one of my missions: bringing Japanese culture into work and sharing it with the world. Come check it out!
Quote
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
⭐️ Ready for the next stage of multi-lingual LMM🌏? 📣 Happy to share our JMMMU🇯🇵, a Japanese MMMU benchmark! For many users, it’s important to accelerate non-English research. JMMMU will accelerate research in Japanese and multi-lingual LMMs! HP: mmmu-japanese-benchmark.github.io/JMMMU/
Image
3.8K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
🎉Happy to share our MM-UPD Leaderboard🏆 huggingface.co/spaces/MM-UPD/ We are now accepting submissions from the community! We look forward to your submissions😀 Many thanks to
@mervenoyann
@clefourrier
for the invitation and for guiding us!
Quote
merve
@mervenoyann
MMUPD is now hosted on @huggingface Hub as a leaderboard 🏆 amazing to see LLaVA-1.6 is outperforming proprietary models on many subtasks 🤩 link in the next one! x.com/mervenoyann/st…
Image
20K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
残念ながら #MIRU2025 は #ACL2025 と日程が被っていていけないのですが、自分が携わった以下の3つの発表があります! これらのワークにて質問等がある場合は自分にDM飛ばしてくれて大丈夫です! よろしくお願いします! また、MIRUに参加される学生に留まらず、漫画の研究、AI
Image
3.8K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
I will be attending #ICLR2024 in person next week and will present UPD at the Workshop on Reliable and Responsible Foundation Models on May 11th. x.com/HuaxiuYaoML/st Please drop by if you can! ☕️" 自分なりに精一杯頑張ります！！時間が合えば是非!
Quote
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
🎉April Fool is the birthday of Unsolvable Problem Detection! UPD examines the VLM’s ability to withhold answers when faced with unsolvable problems. Please enjoy VLMs with unsolvable problems today! paper page: arxiv.org/abs/2403.20331 code: github.com/AtsuMiyai/UPD
6.4K
Atsuyuki Miyai @UTokyo
@AtsuMiyaiAM
We update our #NeurIPS2023, “LoCoOp: Few-Shot Out-of-Distribution Detection via Prompt Learning”.   LoCoOp performs OOD regularization with OOD regions (e.g., background) in ID training images and outperforms existing methods even in a 1-shot setting on ImageNet benchmarks.
Image
2.4K