Gabriel Sarch @GabrielSarch

Efficient, small VLMs for the win!

Liquid AI @liquidai

a week ago

Today, we release LFM2.5-VL-3B, a lightweight vision-language model that reads screens, documents, and the physical world. It handles digital screens across mobile, web, and desktop, grounds objects to coordinates, reads text and charts, and calls tools from either text or image

65 230 2K 293K 1K

My summer project is done! A 20 video, free course on post-training to accompany my book is all on YouTube with slides open for modification & re-use. ~12 hours of content covers the core foundations and some research areas I think will grow in importance. It was a fun time to review all the fundamentals again, as it is clear in the next 1-3 people the amount of people wanting to learn post training will likely 100X again from today, as we have already 100X'ed from two years ago. As AI agents get increasingly capable at coding and discussing these fundamentals (see the code exercises accompanying the book that I am refining with the community) I think developing clear intuitions for how models work and why is one of the most important skills going forward in AI. Still, learning the post-training math is the best way to battle test them. I personally just in this course am starting to master how forward/reverse KL relates to post-training topics. Thanks to all my viewers, and I'm happy to answer questions in the book discord or understand how to better teach the various reward models, on-policy distillation, new RL algorithms, etc. Plus, the book is 50% off right now with the code PBLambert on Manning to celebrate the launch. I'll share the relevant links below. Who's going to make this course for pretraining?

.@danqi_chen, an expert in large language models, has been featured by the Communications of the ACM for her foundational work building small LLMs, which are vital to making modern models available on edge devices such as mobile phones. bit.ly/4vhH5T0

Open weights are great, but open recipes are what make progress reproducible, accessible, and easy to build on. i1 releases a fully open path to strong text-to-image models, including 300+ experiments and lots of training insights. Happy to have played a small role in this.

Zhuang Liu @liuzhuang1234

2 months ago

Can a small academic team build a strong text-to-image model using only public datasets? Introducing i1: a simple, fully open recipe for strong text-to-image models

11 79 441 70K 245

Can a small academic team build a strong text-to-image model using only public datasets? Introducing i1: a simple, fully open recipe for strong text-to-image models

100% agree with @natolambert here. The best recipe work does two things: as science, it explains why something works w/ careful analysis & ablations; for practitioners, makes the findings accessible and usable. We're doing the same for VLMs at Princeton: vero-reasoning.github.io

Nathan Lambert @natolambert

2 months ago

TMax: An open RL recipe for terminal agents I’m very excited to get to share a new RL paper today that I got to have a small part in – a type of paper I suspect we’ll see much more of in the future. The key is that RL research is very different today, in mid-2026, than what most

13 72 659 172K 669

🔵 We’re releasing a major Vero update: • Expanded dataset, models & baselines • Vero-1.6M, expanding Vero-600K to 1.6M samples • Vero-Qwen35-9B and Vero-Qwen35-9B-Base, improving 9B: 74.4 / 73.0 across 30 evals Accepted to #ECCV2026! Check it out: vero-reasoning.github.io

Gabriel Sarch @GabrielSarch

4 months ago

Introducing Vero, the strongest fully open RL recipe for training next-generation visual reasoners. From charts to spatial to open-ended tasks, Vero sets a new bar. • sota 8B VLM across 30 benchmarks • +4.4 avg over four base models (30 evals) • beats prior RL datasets 🧵👇

3 61 302 69K 233

@JangLawrenceK I absolutely love that you called the user Michael Scott. Can’t wait for the Dwight version to run the beet farms B&B business for him. In all seriousness, this is just the kind of realistic benchmark we need for computer use. Very excited about this :)

It’s clear VLMs still struggle across diverse visual inputs. Introducing WorldBench, our new VLM benchmark with visually diverse, realistic, and challenging VQA questions! All questions are annotated *by hand* and checked meticulously for quality. Try it out for yourself!

Harish Krishnakumar @harishkrik

2 months ago

Today’s vision benchmarks suggest VLMs are nearing saturation, but real-world visual understanding is far from solved. Introducing WorldBench: 2,000 hand-written, human-verified VQA questions focused on visual diversity and designed to be challenging for frontier models.

4 17 80 39K 38

Great to see this level of transparency from an industry lab. Detailed recipes and reporting like this are a real asset to the research community and the public!

Hanna Hajishirzi @HannaHajishirzi

3 months ago

MAI-Thinking-1 is out! Excited to share what we are building and how climbing from scratch (no distillation) actually works: simple recipes, rigorous science, self-distillation, patience, and great infra. Check out our tech report has the full story of our RL climbs.

24 126 870 136K 382

I’ll be at CVPR June 3-6 and excited to talk with folks! We’ll be presenting Vero at June 3 workshops (vero-reasoning.github.io): - DataMFM Oral 2:40pm, Room 111 - MMFM Poster, 3pm, 3A - ViScale Poster, 5pm, 506 And be sure to come by the CogVL Workshop cogvl.github.io!

CogVL Workshop at #CVPR2026 is less than a week away! We have an exciting lineup of keynote speakers across Vision, NLP and Cog Sci, and orals/posters on reasoning methods for VLM models. 🕐 June 3, 1 PM 📍 Rooms 610/612 🔗 Schedule: cogvl.github.io

Happy to be recognized as an outstanding reviewer at CVPR!

#CVPR2026 @CVPR

3 months ago

We are grateful to all of the 17,491 reviewers who helped make #CVPR2026 possible. We are especially pleased to recognize the following Outstanding Reviewers, whose high-quality reviews (as judged by their Area Chairs) placed them among the top 5% of reviewers.

keyboard_arrow_left Previous keyboard_arrow_right Next

5 43 226 100K 31

Really excited about this new work scaling VLMs to 100+ turns for playing Mario! Games are a great testbed for long-horizon agentic tasks, and this setting makes it clear the default short-horizon tricks for VLM training often fall short. Check out the thread and paper!

Chengshuai Shi @chengshuai_shi

4 months ago

🔥 Excited to share our new paper: 🚀 Odysseus: Scaling VLMs to 100+ Turn Decision-Making in Games via Reinforcement Learning 🎮 We study how to make RL stable and effective for training VLM agents in long-horizon, visually grounded environments — using the video game Super

5 48 248 104K 171

AI agents can work pretty well on the web now for short tasks. I wanted to know: could they go longer, on harder tasks? Can an agent plan 2 weddings in different cities and a honeymoon within the same month, or find the most suitable culinary arts school across the US for my post PhD plans? We are releasing Odysseys: a benchmark of 200 long-horizon web agent tasks evaluated on the live internet. All our tasks are inspired from real human data and many take hours to complete. The best frontier model we tested (Claude Opus 4.6) reaches only 44.5% perfect-task success, leaving substantial room for improvement. I donated a couple of my own automation wishes to this benchmark. My favorite contribution was to “Rank the top 10 ACL + Meniscus surgeons in the area” - as this took me a decent amount of time to do myself when I got my own knee fixed. GPT 5.5 was able to do this for me with 4 dollars and 30 minutes!

@ysu_nlp @NeoCognition Congratulations @ysu_nlp ! Excited to see what you build!

Hi all. I just put up my course materials for my Multimodal LLM Agents course at @UUtah online so it can be used by the broader community. I really tried to tie together LLMs with RL and CV to give students the full picture of how MLLMs can be incorporated into embodied agents.

Must-read research of the week ▪️ Neural Computers ▪️ The Illusion of Stochasticity in LLMs ▪️ Learning is Forgetting: LLM Training as Lossy Compression ▪️ A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens ▪️ INSPATIO-WORLD: A Real-Time 4D World Simulator via Spatiotemporal Autoregressive Modeling ▪️ Vero: An Open RL Recipe for General Visual Reasoning ▪️ RAGEN-2: Reasoning Collapse in Agentic RL ▪️ TriAttention: Efficient Long Reasoning with Trigonometric KV Compression ▪️ In-Place Test-Time Training ▪️ Fast Spatial Memory with Elastic Test-Time Training ▪️ Gym-Anything: Turn any Software into an Agent Environment ▪️ SkillClaw: Let Skills Evolve Collectively with Agentic Evolver ▪️ PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing Find all the links and other important AI news of the week here: turingpost.com/p/fod148

aimodels.fyi/papers/arxiv/v… > be me > open source researcher > everyone making amazing vision AI but won't share how > "just trust me bro it works" > decide to build our own and document everything > call it Vero > actually works > tfw transparency wins > recipe.pdf

My biggest takeaway from this project: even in an academic setting, with the right people and dedication, we can catch up with part of what top industry teams have achieved. Vero's final training run only takes 8 GPUs. Vero only addresses the RL reasoning stage, but it proves that a lot is still very much possible in academia!

Gabriel Sarch @GabrielSarch

4 months ago

Introducing Vero, the strongest fully open RL recipe for training next-generation visual reasoners. From charts to spatial to open-ended tasks, Vero sets a new bar. • sota 8B VLM across 30 benchmarks • +4.4 avg over four base models (30 evals) • beats prior RL datasets 🧵👇

3 61 302 69K 233