Yuan Yin

AI Researcher @valeoai🦋 | Prev @mlia_isir @Sorbonne_Univ_ | He/him

[Ellington Kirby @EllingtonKirby]

1/🧵 Q: Can we have both a simple and SOTA architecture in autonomous driving?
R: Yes! 😍
Introducing Driving on Registers (DrivoR):
a pure Transformer backbone that achieves SOTA results in NAVSIM v1 / v2 and closed-loop HUGSIM evaluation.
Here is how 👇

[Yuan Yin @yuanyinnn]

Quick clarification: we first shared the Outstanding Paper Award thinking it was the top prize — and it turns out IPA did win Best Paper! Absolutely thrilled! 🎉

[Yuan Yin @yuanyinnn]

1/Serve your PEFT with a fresh IPA!🍺
Finetuning large models is cheaper thanks to LoRA, but is its random init optimal?🤔
Meet IPA: a feature-aware alternative to random projections
#NeurIPS2025 WS #CCFM Oral+Best Paper
Work w/@shawshank_v @tuan_hung_vu @abursuc @quobbe
🧵

[Yuan Yin @yuanyinnn]

8/IPA is just a first step. We believe that a deeper understanding of the feature space is important to unlocking better model adaptation.💭

To try it out and find more details👇:

Arxiv: https://t.co/7Fv6GsIMJN
Code: https://t.co/OavQQ9X6wz

[Yuan Yin @yuanyinnn]

1/Serve your PEFT with a fresh IPA!🍺
Finetuning large models is cheaper thanks to LoRA, but is its random init optimal?🤔
Meet IPA: a feature-aware alternative to random projections
#NeurIPS2025 WS #CCFM Oral+Best Paper
Work w/@shawshank_v @tuan_hung_vu @abursuc @quobbe
🧵

[Yuan Yin @yuanyinnn]

7/Across benchmarks, IPA consistently outperforms standard LoRA and DoRA.
📊Commonsense Reasoning: +1.5 points avg accuracy.
🖼️VTAB-1k (Vision): +2.3 points avg accuracy.
It is also robust at very low ranks (e.g., r=8) where standard LoRA fails.

[elvis @omarsar0]

Learning without training

Google researchers explore the implicit dynamics of in-context learning.

"Implicit weight updates from ICL mirror the effect of actual fine-tuning on the same data."

This one is more technical but much needed.

The findings:

[valeo.ai @valeoai]

🚗 Ever wondered if an AI model could learn to drive just by watching YouTube? 🎥👀

We trained a 1.2B parameter model on 1,800+ hours of raw driving videos.

No labels. No maps. Just pure observation.

And it works! 🤯

🧵👇 [1/10]

[valeo.ai @valeoai]

📚🔍Excited to share our work at #NeurIPS2024! Dive into representation learning, optimization, explainability, VLMs & LLMs, and more.

Check out our blog post to know more about our 7 papers:
https://t.co/gaLDkkYsQe

🧵👇

[Yuan Yin @yuanyinnn]

Trying to move to 🦋. Same id as here😎.

[Victor Letzelter @ ICML 2026 @VLetzelter]

Working on ill-posed machine learning tasks, interested in multi-heads neural networks and data #uncertainty quantification ?

Sharing here our latest research, which will be presented at @NeurIPSConf in December.

[MLIA @mlia_isir]

"GEPS: Boosting Generalization in Parametric PDE Neural Solvers through Adaptive Conditioning"
by @ArmandKassai, @JorgeMifsut, @yuanyinnn, Jean Noël Vittaut, Patrick Gallinari
accepted as conference paper at #NeurIPS2024 !
➡️ https://t.co/ztSKO05ogo
🖥️ https://t.co/Pz4YSrEYLw

[Konstantin Klemmer @kklmmr]

*cries in confirmed reviewer*

[Yuan Yin @yuanyinnn]

@LaureSoulier Merci beaucoup Laure !😉

[LCP @LCP]

🔵🏛️ #IVG dans la Constitution : le Congrès a très largement approuvé le texte qui lui était soumis, faisant de la France le premier pays au monde à inscrire l'IVG dans sa Constitution.
➡️ Récit du Congrès et détail du scrutin, à lire : https://t.co/6JQE1dnR6G
#DirectAN

[Kosta Derpanis (sabbatical in Zurich) @CSProfKGD]

Supervisor's reaction when trainee explains that their 2 week training run failed and the only evidence they have is the loss plot.