Rui Wang

98 posts

Rui Wang

Rui Wang

@theruiwang

Research Lead @yutori_ai, building computer-use models. Previously, post-training lead of Llama 3 Multimodal @AIatMeta, computer vision @FAIR.

Katılım Eylül 2016
166 Takip Edilen210 Takipçiler
Rui Wang retweetledi
Yu Su
Yu Su@ysu_nlp·
Great work by the @yutori_ai team!
Dhruv Batra@DhruvBatra_

𝗡𝗮𝘃𝗶𝗴𝗮𝘁𝗼𝗿 𝗻𝟭.𝟱 “𝘀𝗼𝗹𝘃𝗲𝗱” 𝗢𝗻𝗹𝗶𝗻𝗲 𝗠𝗶𝗻𝗱𝟮𝗪𝗲𝗯: 𝟵𝟳.𝟯% 𝘀𝘂𝗰𝗰𝗲𝘀𝘀 𝗿𝗮𝘁𝗲. While some teams self-report, this result is independently evaluated and verified by OSU NLP Group @osunlp and Careerflow Human Data Labs. All benchmarks are transient attempts at measuring progress. Ultimately, what matters is how a model performs when people use it. But there’s a sentiment online that computer-use models aren’t progressing quickly. Not true. In the last year, performance on Online Mind2Web has gone from ~40% success to basically saturated. So what’s next? Most computer-use/browser-use benchmarks are GUI-only. Models (including Navigator n1.5) now support hybrid actions — UI interactions (click, type, scroll) and programmatic actions (e.g., execute JS). Ultimately, we’re headed to a world where computer-use models “agentify” the long-tail of the web. huggingface.co/spaces/osunlp/…

English
0
5
12
3.3K
Rui Wang
Rui Wang@theruiwang·
Online-Mind2Web is "solved". We’ve received officially verified results from the OSU NLP Group: - 97.3% human-verified accuracy, following three independent reviews, additional QA on borderline cases, and final manual verification by the benchmark authors. - 87.9% auto-eval accuracy with WebJudge, with all inputs and outputs from all three evaluation stages submitted for full reproducibility. We thank the benchmark authors @xue_tianci @hhsun1 @ysu_nlp for upholding high academic standards and applying consistent evaluation protocols across official submissions. Their work gives the field a fair and rigorous way to measure progress and compare models directly in the real world. One year ago, SOTA performance on Online-Mind2Web was around 50%. Today, it stands at 97.3%—a significant milestone for computer-use agents. What a year!
Dhruv Batra@DhruvBatra_

𝗡𝗮𝘃𝗶𝗴𝗮𝘁𝗼𝗿 𝗻𝟭.𝟱 “𝘀𝗼𝗹𝘃𝗲𝗱” 𝗢𝗻𝗹𝗶𝗻𝗲 𝗠𝗶𝗻𝗱𝟮𝗪𝗲𝗯: 𝟵𝟳.𝟯% 𝘀𝘂𝗰𝗰𝗲𝘀𝘀 𝗿𝗮𝘁𝗲. While some teams self-report, this result is independently evaluated and verified by OSU NLP Group @osunlp and Careerflow Human Data Labs. All benchmarks are transient attempts at measuring progress. Ultimately, what matters is how a model performs when people use it. But there’s a sentiment online that computer-use models aren’t progressing quickly. Not true. In the last year, performance on Online Mind2Web has gone from ~40% success to basically saturated. So what’s next? Most computer-use/browser-use benchmarks are GUI-only. Models (including Navigator n1.5) now support hybrid actions — UI interactions (click, type, scroll) and programmatic actions (e.g., execute JS). Ultimately, we’re headed to a world where computer-use models “agentify” the long-tail of the web. huggingface.co/spaces/osunlp/…

English
2
5
12
1.3K
Rui Wang
Rui Wang@theruiwang·
@ericjang11 Happy to chat! My robots crash in operating systems instead of into walls 😄
English
0
0
0
115
Eric Jang
Eric Jang@ericjang11·
Lately I've been learning about multimodal models. I'd like to learn how to make them better. Who should I meet? What lectures should I watch? For multimodal folks interested in chatting, I'd be happy to do a "lunch swap" on where I see robotics going in the next 24 months 🥪
English
23
6
225
31.9K
Kevin Chih-Yao Ma
Kevin Chih-Yao Ma@chihyaoma·
Just 2 months after our previous launch, MAI-Image-2.5 preview is out today! Currently, sitting at top3 on Arena. It's a preview :) More to come.
Arena.ai@arena

Exciting news, MAI-Image-2.5 (Preview) from @MicrosoftAI debuts at #3 in the Text-to-Image Arena with a score of 1,254 — a +72 point improvement over MAI-Image-2. A top 5 arena previously held only by @GoogleDeepMind and @OpenAI has a new lab in the mix. Congrats to the @MicrosoftAI team on this accomplishment.

English
5
5
46
6K
Rui Wang
Rui Wang@theruiwang·
n1.5 is now serving production CUA traffic at a FAANG company after extensive evals against models from big labs on both accuracy and efficiency. If you’re currently using Claude / Gemini / GPT as a web agent, it’s worth trying Navigator n1.5 — it will likely reduce cost, improve latency, and deliver stronger performance.
Devi Parikh@deviparikh

We gave some of our partners early access to n1.5 — the most capable computer use model for the web. It is in production at FAANG scale as we speak, replacing a computer use model from a frontier lab. If your product can benefit from web automation — extracting structured data from dynamic webpages, filling forms, completing workflows on the web, testing vibe coded web apps — you should try out @yutori_ai's Navigator n1.5! Save your GPT / Claude / Gemini capacity for something else :)

English
2
5
20
2.8K
Rui Wang
Rui Wang@theruiwang·
A conceptually simple web task can become a “long-horizon” task for today’s computer-use models. If a model can only act through UI primitives, click, type, scroll, everything gets forced into a sequence of low-level operations. Even simple tasks turn into dozens of steps: filling forms field by field, clicking through variants, and executing long sequences of small, incremental interactions. We introduced a hybrid mode: vision + DOM + code. Pure vision generalizes but is unnecessarily slow; pure DOM + code is fast but can be brittle given the current state of the web. The hybrid combines both, taking the fast, programmatic path when possible, while using vision to verify against what’s actually rendered. The result: "long-horizon" tasks collapse into shorter ones—less time, less compute, and more reliable outcomes.
Dhruv Batra@DhruvBatra_

𝐈𝐧𝐭𝐫𝐨𝐝𝐮𝐜𝐢𝐧𝐠 𝐍𝐚𝐯𝐢𝐠𝐚𝐭𝐨𝐫 𝐧𝟏.𝟓 The most capable computer-use model for the web. Pareto-domination: accuracy, latency, cost • SoTA across all benchmarks • +5-10% over GPT 5.5, Opus 4.7, n1 • +25% over Gemini • 2x faster, significantly cheaper Expanded action space • UI actions (like n1) + JavaScript generation & execution

English
2
6
18
118.3K
Dhruv Batra
Dhruv Batra@DhruvBatra_·
𝐈𝐧𝐭𝐫𝐨𝐝𝐮𝐜𝐢𝐧𝐠 𝐍𝐚𝐯𝐢𝐠𝐚𝐭𝐨𝐫 𝐧𝟏.𝟓 The most capable computer-use model for the web. Pareto-domination: accuracy, latency, cost • SoTA across all benchmarks • +5-10% over GPT 5.5, Opus 4.7, n1 • +25% over Gemini • 2x faster, significantly cheaper Expanded action space • UI actions (like n1) + JavaScript generation & execution
Dhruv Batra tweet media
English
15
23
193
152.4K
Rui Wang retweetledi
Devi Parikh
Devi Parikh@deviparikh·
Why 996 when 007 has you covered? Try Yutori Delegate.
English
2
3
23
2.3K
Rui Wang
Rui Wang@theruiwang·
RT @DhruvBatra_: Introducing Yutori Delegate. Why co-work when you can delegate?
English
0
1
0
28
Rui Wang retweetledi
Devi Parikh
Devi Parikh@deviparikh·
This is what we started Yutori for, and today we're launching it! Meet Delegate. Designed for delegation, not chat. It’s proactive, it’s always on, it remembers. It connects to your apps and learns your context — who you are, how you work and what your priorities are. It schedules itself forward, it follows up, it follows through. You don’t need the perfect prompt. Just do a stream of consciousness brain dump of everything on your mind. Delegate it all. And move on. It will break things down. Finish what it can. Come back for input where it needs it. Try it! yutori.com/delegate
Abhishek Das@abhshkdz

We're excited to launch Delegate. An agent you delegate work to and move on with your life.

English
2
6
42
16.3K
Rui Wang retweetledi
Abhishek Das
Abhishek Das@abhshkdz·
We're excited to launch Delegate. An agent you delegate work to and move on with your life.
English
30
31
187
66.7K
Rui Wang retweetledi
Devi Parikh
Devi Parikh@deviparikh·
Yutori at this morning's @Nasdaq bell ringing :) Thank you @notablecap and partners for including us in the Prosumer AI 40 list!
Devi Parikh tweet mediaDevi Parikh tweet media
English
5
5
60
5.8K
Rui Wang retweetledi
Abhishek Das
Abhishek Das@abhshkdz·
We just shipped the biggest update to Scouts since launch (and yes, we know what day it is). Scouts used to be just for monitoring. Now they act. Scouts is now a general-purpose task execution engine for the web. Tell it what you need done, and it does it: across any website, behind any login, connected to your apps. 🧵
GIF
English
4
10
82
20.2K

Keşfet

@yutori_ai @xue_tianci @hhsun1 @ysu_nlp @ericjang11 @chihyaoma @DhruvBatra_ @elonmusk