Yu Su
@ysu_nlp
Followers
12K
Following
4K
Media
137
Statuses
2K
something new | prof. @osunlp | sloan fellow | intelligence and agents | author of Mind2Web, SeeAct, MMMU, HippoRAG, BioCLIP, UGround.
Columbus, OH
Joined March 2013
Computer Use: Modern Moravec's Paradox A new blog post arguing why computer-use agents may be the biggest opportunity and challenge for AGI. https://t.co/6fZfTdx710 Table of Contents > Moravec’s Paradox > Moravec's Paradox in 2025 > Computer use may be the biggest opportunity
9
65
207
I don't necessarily agree with everything Microsoft is doing in the AI space, but man, Satya is such a fantastic CEO: sharp, grounded, long-termist, and still super easy to be around with. I should have held my MSFT for longer.
.@satyanadella gave me and @dylan522p an exclusive tour of Fairwater 2, the most powerful AI datacenter in the world. We then chatted through Satya's vision for Microsoft in a world with AGI. 0:00:00 - Fairwater 2 0:04:15 - Business models for AGI 0:13:42 - Copilot 0:20:56 -
1
2
17
Excited for this week's AI Agent Frontier Seminar! We're thrilled to host Dr. Huan Sun (@hhsun1 ) from @OhioState. She'll discuss a critical topic: "Advancing the Capability and Safety of Computer-Use Agents Together." 🗓️ Friday, Nov 14 ⏰ 9 AM PT / 12 PM ET All are welcome!
0
2
13
In my ICLR SAC batch (225 papers), only 18 paper (8%) have an avg initial rating >= 6 (borderline accept). Maybe my batch is not the most representative, or maybe it's reviewer fatigue from the exploding # of submissions. Silver lining is, don't give up on rebuttal.
11
12
238
🚀 Worried about faculty openings? Ohio State @OhioState is to hire 100 new faculty with AI expertise over the next five years! 🤖🎓 The new hires will join one of three AI Faculty Cohorts: 🧠 Foundational AI — Elevating the theoretical, mathematical, and algorithmic
8
42
211
Super fun to serve on LEAP's academic advisory board and observe the forecasting. If the development of AI can live up to the median forecasts, it will already be a big deal.
Today, we are launching the most rigorous ongoing source of expert forecasts on the future of AI: the Longitudinal Expert AI Panel (LEAP). We’ve assembled a panel of 339 top experts across computer science, AI industry, economics, and AI policy. Roughly every month—for the next
0
1
15
⛵Marin 32B Base (mantis) is done training! It is the best open-source base model (beating OLMo 2 32B Base) and it’s even close to the best comparably-sized open-weight base models, Gemma 3 27B PT and Qwen 2.5 32B Base. Ranking across 19 benchmarks:
19
83
558
We just built and released the largest dataset for supervised fine-tuning of agentic LMs, 1.27M trajectories (~36B tokens)! Up until now, large-scale SFT for agents is rare - not for lack of data, but because of fragmentation across heterogeneous formats, tools, and interfaces.
arxiv.org
Public research results on large-scale supervised finetuning of AI agents remain relatively rare, since the collection of agent training data presents unique challenges. In this work, we argue...
27
173
1K
📢 As AI becomes increasingly explored for research idea generation, how can we rigorously evaluate the ideas it generates before committing time and resources to them? We introduce ScholarEval, a literature grounded framework for research idea evaluation across disciplines 👇!
4
42
141
The GOAT
2
4
117
Genuine question: how is @OpenAI Atlas different from @perplexity_ai Comet? at first glance the set of AI features looks quite similar
1
0
10
I am on the faculty job market this year! I am seeking tenure-track faculty positions to drive my research agenda on rigorous AI evaluation for science and policy. I am applying broadly across disciplines, and would be grateful to hear of relevant positions. Materials: 🧵
10
71
420
I've given 30+ talks on agents in the past 2 years, and I always end my talk with this slide. We are just at the dawn of a long journey on agents. General agents need the same broad set of cognitive competencies as humans and more. It doesn't necessarily have to be constructed
> @karpathy is right on the limitations of current agents > 2025 is year 0 of the decade of agents: agents started to bring significant marginal value in narrow domains (eg, coding, customer service), but still far from general human-level competency > agents need a wholesale
6
13
131
R.I.P.
Prof. Chen Ning Yang, a world-renowned physicist, Nobel Laureate in Physics, Academician of the Chinese Academy of Sciences, Professor at Tsinghua University, and Honorary Director of the Institute for Advanced Study at Tsinghua University, passed away in Beijing due to illness
0
0
14
> @karpathy is right on the limitations of current agents > 2025 is year 0 of the decade of agents: agents started to bring significant marginal value in narrow domains (eg, coding, customer service), but still far from general human-level competency > agents need a wholesale
BREAKING: Andrej Karpathy calls out Sam Altman Altman: >"We are now confident we know how to build AGI" >"2025 is the year of AI agents" Karpathy: >"I was triggered by that over-prediction" >"More accurately, it's the decade of agents" >"There's SO much work to be done" Hmm
3
5
69
Excited to introduce a new agent learning paradigm called Early Experience, as a reward-free mid-training stage for large-scale agent training. A fantastic collaboration between Meta Superintelligence Lab and @osunlp led by the amazing @KaiZhang_CS. Built on insights from our
arxiv.org
Language agents based on large language models (LLMs) have demonstrated great promise in automating web-based tasks. Recent work has shown that incorporating advanced planning algorithms, e.g.,...
🌀Agent Learning via Early Experience🌀 📝: https://t.co/VsqQHTTrBN - SFT for agents is sparse; RL on long-horizons is hard We provide new mid-training signals that work: 1) Implicit next state world modeling task 2) Self-reflection on alternate states - Strong improvements over
5
29
175
Introducing early experience: using future states resulting from agent’s own action as scalable supervision to train itself - without reward🧠! 1️⃣Reward-free: can train directly in real-world environments. 2️⃣Better RL warm-start: when continued with RL, leads to higher final
🌀Agent Learning via Early Experience🌀 📝: https://t.co/VsqQHTTrBN - SFT for agents is sparse; RL on long-horizons is hard We provide new mid-training signals that work: 1) Implicit next state world modeling task 2) Self-reflection on alternate states - Strong improvements over
2
28
109
Really excited to announce AgentX–AgentBeats Competition 🚀 💰 $1 Million+ in prizes, cloud credits, and API resources, a global challenge hosted by @BerkeleyRDI , building on the Agentic AI MOOC community of 32K+ learners, bringing together builders, researchers, engineers, and
11
28
84
The most comprehensive agent evaluation to date
📣New paper: Rigorous AI agent evaluation is much harder than it seems. For the last year, we have been working on infrastructure for fair agent evaluations on challenging benchmarks. Today, we release a paper that condenses our insights from 20,000+ agent rollouts on 9
1
13
131
New @GoogleResearch paper shows agents learn software skills by watching tutorials, converting them into action steps, and boosting task performance. So converts free videos into reliable supervision at scale. A vision model, inverse dynamics, predicts the action between 2
14
82
441