
Research Scientist @GoogleDeepMind - LLM post-training - prev: @Google @MSFTResearch
Montréal, Québec
Joined March 2011
- Tweets429
- Following359
- Followers2.3K
- Likes2.1K
After nearly two years, I’ve decided to leave GDM.
It was an incredible journey working on post-training, distillation and agentic reasoning for Gemini and Gemma. But the people made it special.
Feeling lucky to have worked alongside so many brilliant minds.
13
127
5,421
special shout-out to Doina, Yinlam, @AleksandraFaust @kazemi_sm @BahareFatemi @sanketvmehta @vqctran @_angie_chen @ankit_s_anand @agarwl_ @johanferret @sarah_perrin_ @robdadashi @OlivierBachem @charlinelelan and many more. I learned so much from you all.
8
584
Arian Hosseini retweeted
Want to train your SWE agents 2× faster 🏎️💨?
We introduce KV-streams, which, rather than flushing the KV cache on agentic compaction, we preserve it and keep generating. This avoids the re-prefill cost on both the inference engine and the trainer.
Unlike other works that assume a fixed compaction strategy, we show that KV-streams can work with any compaction method in agentic tasks.
Using what we learn from experiments in text-based games. We scale KV-streams to SWE tasks, where we match the performance of re-prefill compaction and full context in half the time 🚀
arxiv.org/pdf/2609.35750
🧵
So proud of my Montrealer 🇨🇦 collaborators
@BahareFatemi , @kazemi_sm , and our whole team for the contributions to the agentic capabilities behind this release.
Huge milestone today! 🚀
I’m in Seoul for ICML 🇰🇷 I’ll be at the Google DeepMind booth. Our team in Mountain View is hiring researchers. Come by if you’re interested or just want to chat 👋🏼
Our team at Google DeepMind is hiring a Research Scientist in Mountain View!
Focus is on bridging foundational research & practical GenAI deployments.
✅ PhD in CS/ML
✅ LLMs & diffusion expertise
✅ Top-tier ML pubs
✅ Python/JAX/PyTorch
Apply : google.com/about/careers/app…
Our team at GDM is hiring a Student Researcher (SR) next year 🧠
If you’re a PhD student working on LLMs please apply. I’d love to hear from you.
Please fill out this form: forms.gle/bxTEkrDPacn6jSxi6
The new Gemini is out and the reasoning capabilities are next level 🧠 fun one to be a part of. Congrats to the whole team!
1
6
3,438
Come by ScaleR #COLM2025. Nishad and I will be presenting this work!
Want to know how to optimally scale test-time compute with generative verifiers?
Come to our poster at #COLM2025 🇨🇦.
hmu if you want to chat about LLM reasoning and agents! :)
Poster session 1 (Tuesday Morning 11 AM - 1 PM), Poster number: 24.
Arian Hosseini retweeted
Introducing linear scaling of reasoning:
𝐓𝐡𝐞 𝐌𝐚𝐫𝐤𝐨𝐯𝐢𝐚𝐧 𝐓𝐡𝐢𝐧𝐤𝐞𝐫
Reformulate RL so thinking scales 𝐎(𝐧) 𝐜𝐨𝐦𝐩𝐮𝐭𝐞, not O(n^2), with O(1) 𝐦𝐞𝐦𝐨𝐫𝐲, architecture-agnostic.
Train R1-1.5B into a markovian thinker with 96K thought budget, ~2X accuracy 🧵
Come by # 82 at #COLM2025
Arian Hosseini retweeted
Our new work probes LLMs’ multi-turn reasoning and planning capabalities via five challenging games. Check it out!
LLMs are great at single-shot problems, but in the era of experience, interactive environments are key 🔑
Introducing * Multi-Turn Puzzles (MTP) * , a new benchmark to test multi-turn reasoning and strategizing
🔗 Paper: huggingface.co/papers/2508.1…
🫙Data: huggingface.co/datasets/aria…
LLMs are great at single-shot problems, but in the era of experience, interactive environments are key 🔑
Introducing * Multi-Turn Puzzles (MTP) * , a new benchmark to test multi-turn reasoning and strategizing
🔗 Paper: huggingface.co/papers/2508.1…
🫙Data: huggingface.co/datasets/aria…
Our analysis shows stronger models don't just guess better. they build coherent plans, ask complex questions, and re-evaluate their own conclusions based on new info.
Weaker models jump to conclusions and ask less or questions. It's simple Q&A vs strategic dialogue.
7/n
1
4
354
MTP provides a robust platform to diagnose weaknesses and improve multi-turn capabilities.
Enjoyed working with @BadolaKartikeya @kazemi_sm @BahareFatemi Jon Simon and the rest of the team @GoogleDeepMind .
Details: huggingface.co/papers/2508.1…
Data & more: huggingface.co/datasets/aria…
4
342














