
Research scientist at @NVIDIAAI | PhD at @ids_ou | Research intern @SonyAI_global, @NVIDIAAI | Interested in fairness in Vision and Language
Taipei, Taiwan
Joined May 2022
- Tweets62
- Following504
- Followers551
- Likes1.6K
Yusuke Hirota retweeted
GPT-6 Astraに「Qwen3.5-0.8Bを数学に強くするように重みを書き換えて」と頼んだところ、全体のたった0.016%の重み変更で数学のベンチマークでは6倍・他のベンチマークでは最大3倍ほど性能が上がりました。
MITに来てからは最初の成果です!
rikiyat.github.io/blog/weigh…
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Yusuke Hirota retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
Does giving a multimodal model tools make it worse at refusing harmful requests?
New work from NVIDIA, accepted at NeurIPS 2026, says yes for every model it tested.
Refusal failures rise by up to 68.7% relative, and by 17.7% on average.
The drop appears in Claude Opus 4.6 and 4.7, Gemini Agentic Vision, Qwen3.5-122B-A10B, and agent-tuned open models across MM-SafetyBench, HoliSafe, and VLSBench.
The authors trace it to two causes.
Tool outputs fill the context and bury the original request's harmful intent. The model also shifts its attention to describing what the tools returned instead of making the safety decision.
Re-inserting the original request and image right before the final response restores part of the lost refusals.
If your safety evals run only in plain chat, they may overstate how safe your agent is.
Paper: arxiv.org/abs/2610.03938
Chat with Paper: academy.dair.ai/papers/mllms…
Yusuke Hirota retweeted
NVIDIA Research Taiwanチームで2027 Internshipの応募が公開されました!
僕は以下のトピックで募集しています🙌
✅ Safety for Embodied Agents
✅ Video Understanding
✅ Agentic VLMs
....
興味ある方はメールかDMください!
rhachiuma [at] nvidia.com
nvidia.wd5.myworkdayjobs.com…
Yusuke Hirota retweeted
🚀 Releasing ✨AXPO✨ an RL method to lift agentic reasoning models past their next scaling tier.
Be it math, perception, or search, AXPO fixes the structural blind spot 'just add tools' recipes leave untouched.
8B beats 4x larger 32B baseline on Pass@4.
from NVIDIA 🧵 (1/7)
Yusuke Hirota retweeted
今年のFull-time position (新卒・中途)まだまだ募集してますのでお気軽にお声がけください!!
NVIDIA Research TaiwanチームでFull-time positionの応募が公開されました!
New college gradsと書いてありますが、博士新卒じゃなくてもOKです👌
興味ある方は直接応募いただくかDMください!
nvidia.eightfold.ai/careers/…
Yusuke Hirota retweeted
As AI becomes more integrated into our lives, building Trustworthy AI isn't just a goal, it's a necessity. Exciting to see NVIDIA Research Taiwan tackling these safety challenges. If you want to make an impact, don't miss this! Yusuke is incredibly passionate about this space.
🚀NVIDIA Research Taiwan @NVIDIAAI で研究インターンを募集しています!
研究トピックは、Trustworthy AI、Safety in Agentic Tasks、Content Moderation などです。
興味のある方は、DM やメールで気軽にご連絡ください!
🚀NVIDIA Research Taiwan @NVIDIAAI で研究インターンを募集しています!
研究トピックは、Trustworthy AI、Safety in Agentic Tasks、Content Moderation などです。
興味のある方は、DM やメールで気軽にご連絡ください!
Yusuke Hirota retweeted
NVIDIA Research TaiwanチームでFull-time positionの応募が公開されました!
New college gradsと書いてありますが、博士新卒じゃなくてもOKです👌
興味ある方は直接応募いただくかDMください!
nvidia.eightfold.ai/careers/…
Yusuke Hirota retweeted
Replying to @cvl_osaka @ICCVConference
[Poster 2 #707] @dahlian0+, AgroBench: Vision-Language Model Benchmark in Agriculture
dahlian00.github.io/AgroBenc…
[Poster 2 #801] @hirota_yusuke+, Bias in Gender Bias Benchmarks: How Confounding Features Distort Evaluation
Our team at #NVIDIA Research Taiwan is hiring interns! (Topic: Fairness). If you’re attending #ICCV2025, feel free to chat in person!
I’ll also be presenting our poster:
“Bias in Gender Bias Benchmarks”
🗓️ Oct 21st at Exhibit Hall I #801 (afternoon)
🔗 arxiv.org/abs/2509.07596
Yusuke Hirota retweeted
#MIRU2025 では、5件の共著の発表があります!
僕も現地にいるので、NVIDIA Taiwanでの来年以降のインターンやFull-timeなど興味ある方もお気軽にお声掛けください!
IT2-03 (Masashi Hatano), IS2-056 (@_ko4ro_ ), IS3-072 (@_ko4ro_ ), IS3-150 (@017_ishikawa ), IS3-180 (@ry0fujii )
Yusuke Hirota retweeted
Replying to @nvidia
@nvidia research will present a few NLP works at @aclmeeting ACL 2025 at Austria 🇦🇹
NEKO: Cross-Modality Post-Recognition Error Correction with Tasks-Guided Mixture-of-Experts LM
Industry Session Oral, Hall L 11:00 to 12:30 CET, Monday
aclanthology.org/2025.acl-in…
(@yentinglin56 @ZhehuaiC @PiotrZelasko et al. w/ NVR and NeMo)
LOTUS: A Leaderboard for Detailed Image Captioning from Quality to Societal Bias and User Preferences
Industry Session Poster, Hall 4X 16:00 to 17:30 CET, Tuesday
aclanthology.org/2025.acl-in…
(@hirota_yusuke @Boyiliee @RHachiuma et al. w/ NVR )
Feel free to come to Frank Wang and us in-person, if you are interested in the related topics. 🙂
Don’t miss it! 🚨 The submission deadline for the STREAM Workshop at #ICCV2025 @ICCVConference is June 30!
Welcoming full papers & short abstracts on Trustworthy AI, Fairness, Explainability, and more.
🔗 stream-workshop.github.io/st…
Glad to have a paper on “bias in gender bias benchmarks” accepted at #ICCV2025!
Also excited to announce the STREAM Workshop at #ICCV2025! — submissions open for full papers & short abstracts.
Topics: Trustworthy, Fairness, Explainable AI, and more.
🔗 stream-workshop.github.io/st…
Yusuke Hirota retweeted
⏳ Don’t miss out! Submit your work to the STREAM Workshop @ICCVConference by June 30.
📄 Call for Papers: stream-workshop.github.io/st…
📥 Submission site: cmt3.research.microsoft.com/…
Glad to have a paper on “bias in gender bias benchmarks” accepted at #ICCV2025!
Also excited to announce the STREAM Workshop at #ICCV2025! — submissions open for full papers & short abstracts.
Topics: Trustworthy, Fairness, Explainable AI, and more.
🔗 stream-workshop.github.io/st…
Yusuke Hirota retweeted
Yusuke joined our research team from this April!🇹🇼
If you are interested in the fairness topics, feel free to send a message to him!
Also, I’ll be at #ICLR2025 🇸🇬 to present our paper "SANER: Annotation-free Societal Attribute Neutralizer for Debiasing CLIP" (Poster Session 6, Saturday 3pm)
If you’re interested in fairness or curious about research internships at NVIDIA Taiwan, let’s chat!

















