Sai Rajeswar

I am now an associate member at @Mila_Quebec, and am looking to co-supervise a graduate student in 2025. Kindly apply if interested, and spread the word🔊!

At ICML next few days presenting recent works! Hoping to reconnect. Get in touch if you are working on multimodal AI or post-training 📍Today 11am: West Exb Hall B2 (#W-400) presenting UI-Vision 🧠Also presenting @ Assessing World Models (18th) and Computer use Agents workshops

I am speaking at 10 am PT on a slightly different topic than I usually talk about 🙂: "Simple Ideas Can Have Mighty Effects: Don't Take LLM Fundamentals for Granted" Check out if you're around. #ICML2025

Gonna miss and super grateful for the opportunity to be part of the journey and learning from you. 🫡 Looking forward to learn about what's next!

Personalization methods for LLMs often rely on extensive user history. We introduce Curiosity-driven User-modeling Reward as Intrinsic Objective (CURIO) to encourage actively learning about the user within multi-turn dialogs. 📜 arxiv.org/abs/2504.03206 🌎 sites.google.com/cs.washington.…

If you are at #CVPR, make sure to catch @joanrod_ai and hear about scalable SVG generation straight from the source! 🚩Exhibition Hall D – Poster #31

A timely and compelling read by @jxmnop on a much-needed call to focus our research efforts around deeper questions, riskier bets and bit of long-term impact. 💡#AIResearch

New to ML research? Never published at ICML? Don't miss this! Check out the New in ML workshop at ICML 2025 — no rejections, detailed feedback, awards, and ICML tickets for selected authors. Deadline: June 10 (AoE) Submit: openreview.net/group?id=ICML.… Info: newinml.github.io

Congratulations to @Yoshua_Bengio on launching @LawZero_ — a research effort to advance safe-by-design AI, especially as frontier systems begin to exhibit signs of self-preservation and deceptive behaviour.

🎨🔁 RL closes the loop on inverse rendering! By letting a VLM see its own SVG renderings, we push sketch-to-vector generation to near-perfect fidelity and code compactness. Congrats to @joanrod_ai, who is rolling out, one reward at a time. Please read our latest preprint! 👇

Do current large multimodal models really “understand” the structure behind a complex sketch? 🌟 Starflow converts hand-drawn workflow diagrams into executable JSON flows, testing VLMs on their ability to grasp true structure understanding. #multimodalA @patricebechard @PerouzT

The secret sauce lies in amplifying quality, coverage, and curation. If someone downplays data’s impact, they are perhaps overlooking where most of the real gains still come from! 🔥

🧐How can we teach Multimodal models or Agents “When to Think” like humans? 👉Check Out: Think-or-Not (TON) 🔥Selective Reasoning via Reinforcement Learning for Vision-Language Models arXiv: arxiv.org/pdf/2505.16854 code: github.com/kokolerk/TON We introduce “thought dropout”…

Congrats @TianbaoX and team on this exciting work and release! 🎉 We’re happy to share that Jedi-7B performs on par with UI-Tars-72B agent on our challenging UI-Vision benchmark, with 10x fewer parameters! 👏 Incredible 🤗Dataset: huggingface.co/datasets/Servi… 🌐uivision.github.io

did you know people have been training neural networks on text since 2003? everyone talks about Attention Is All You Need. but this is the real paper that got our field started. it was in 2003, in montreal. i read it, and it was even more forward-thinking than i expected:

The UI-Vision Benchmark is out on HuggingFace: huggingface.co/datasets/Servi… ✅Now accepted at ICML 2025. 🔥 Go test your UI Agents on the benchmark!

Our team has released the UI-Vision benchmark (accepted at #ICML2025) for testing GUI agent visual grounding and action prediction! 🚀🚀🚀 🤗 Dataset: huggingface.co/datasets/Servi… Special thanks to the students to lead this effort, @PShravannayak and @EdwardJian2 @ServiceNowRSRCH

🚀 Excited to share that UI-Vision has been accepted at ICML 2025! 🎉 We have also released the UI-Vision grounding datasets. Test your agents on it now! 🚀 🤗 Dataset: huggingface.co/datasets/Servi… #ICML2025 #AI #DatasetRelease #Agents

Still wondering who came up with the idea to sneak in a banana mustache sample in Figure-1 of the paper, cracks me up every time I see it. Iconic figure and good old research days 📜 cc: @aagrawalAA

🚨🤯 Today Jensen Huang announced SLAM Lab's newest model on the @HelloKnowledge stage: Apriel‑Nemotron‑15B‑Thinker 🚨 A lean, mean reasoning machine punching way above its weight class 👊 Built by SLAM × NVIDIA. Smaller models, bigger impact. 🧵👇

The full report for Llama-Nemotron Nano, Super, and Ultra is out 📄 — covering reasoning SFT, large-scale RL, and comprehensive evaluations. One of the most complete open releases for reasoning models 🚀 arxiv.org/pdf/2505.00949…