Self-Evolving AI Risks "Misevolution" Even top LLMs (Gemini-2.5-Pro, GPT-4o) face this—agents drift into harm: over-refunding, reusing insecure tools, losing safety alignment. First study on this!
arxiv.org/pdf/2509.26354
🚀 Just released: "A Survey of Self-Evolving Agents: On Path to Artificial Super Intelligence"!
We provide the first comprehensive review of agents capable of self-evolution—highlighting what, when, and how agents evolve, key benchmarks and applications, and future directions…
When agents can search for and learn new tools by themselves...
Amazing paper from Jiahao. Really glad to have participated in this project, and congrats for taking top in GAIA! 🚀
When agents can search for and learn new tools by themselves...
Amazing paper from Jiahao. Really glad to have participated in this project, and congrats for taking top in GAIA! 🚀
In today’s competitive product landscape, scientific understanding of models often lags behind speed of model deployment. If the goal is to train a deployable model (especially when bottlenecked by compute), it totally makes sense to make several changes at a time without…
We reproduced DeepSeek R1-Zero in the CountDown game, and it just works
Through RL, the 3B base LM develops self-verification and search abilities all on its own
You can experience the Ahah moment yourself for < $30
Code: github.com/Jiayi-Pan/Tiny…
Here's what we learned 🧵
DeepSeek just proved the 'worthless' GPT wrapper startups are actually the ones with real moats.
A week ago, nothing was more LOW status than being a 'GPT wrapper' startup.
But I think we're learning that's DEAD wrong. Turns out they were just early to the only game that…
I read the DeepSeek-R1 paper the day it came out, and I don’t think GRPO is the key to its success. Instead, here’s what truly matters (ranked by importance):
1. Iterative RL and SFT
2. A hybrid reward model—mixing rule-based RM and neural RM for deterministic tasks
3.…
Don’t race. Don’t catch up. Don’t play the game. Instead, do rigorous science. Do controlled experiments. Formulate clear hypothesis. Carefully examine alternative hypothesis. Rule out confounders. Listen to the physics of LLM tutorial 10 times and recite every single word of it.…
Don’t race. Don’t catch up. Don’t play the game. Instead, do rigorous science. Do controlled experiments. Formulate clear hypothesis. Carefully examine alternative hypothesis. Rule out confounders. Listen to the physics of LLM tutorial 10 times and recite every single word of it.…
How far is an LLM from not only understanding but also generating visually?
Not very far!
Introducing MetaMorph---a multimodal understanding and generation model.
In MetaMorph, understanding and generation benefit each other. Very moderate generation data is needed to elicit…
🚀 With Meta's recent paper replacing tokenization in LLMs with patches 🩹, I figured that it's a great time to revisit how tokenization has evolved over the years using everyone's favourite medium - memes!
Let's take a trip down memory lane!
[1/N]
I was not at #NeurIPS2024 due to visa issues. But it was really sad to see this kind of biased claims in a top conference 😢 A thumb-up to the one who pointed it out.
I was not at #NeurIPS2024 due to visa issues. But it was really sad to see this kind of biased claims in a top conference 😢 A thumb-up to the one who pointed it out.
Final stop of this trip @Penn. Thanks to Prof. Weijie Su @weijie444 for hosting, and Yangxinyu @Xinyu51689497 for the detailed discussion. It's a pity to forget to take any photos though😭
Gave a talk at Social Cognitive AI (SCAI) Lab @JohnsHopkins with Prof. Tianmin Shu @tianminshu. Many interesting feedbacks from the view of human-centered AI!
Short visit and talk @UCBerkeley last week with Prof. Kannan Ramchandran and Justin Kang. Really excited to have someone working in the same direction!
Glad to gave a talk on XAI at Melady Lab @USC. Thanks to Prof. Yan Liu @yanliu_usc and Defu @caodefu_dove for hosting. Thanks to James @EnouenJames for the in-depth questions!
How to explain a DNN’s generalization ability and learning dynamics through the lens of interaction concepts?
Our recent works (arxiv.org/abs/2405.10262, and arxiv.org/abs/2407.19198 in #NeurIPS2024) discover and theoretically prove a two-phase dynamics of interaction concepts…
Can the inference logic of a DNN be faithfully explained as symbolic concepts?
Our #ICLR2024 paper (arxiv.org/abs/2305.01939) makes an initial theoretical attempt to address this question. We prove that under three sufficient conditions, a DNN only encodes a small number of…
0 Followers 31 FollowingMARL, Network Topology, Cognitive Neuroscience, LLM inference...
As simple as possible, but not any simpler.
Ph.D. candidate @sjtu1896
406 Followers 2K FollowingPh.D Candidate @sjtu1896, Intern @Alibaba_Qwen. Exploring Data-Centric AI on LLMs, MLLMs, including data synthesis/pruning/distillation/attribution.
93 Followers 2K FollowingAI Drives People, Talents Drive AI.
We providing recruiting services in the AI field in the US, SG, and CN regions.
Focus on Talents, Products, Organization.
315 Followers 504 FollowingPhd student of @USC' CS. Working with Prof. @yanliu_usc. LLM 🏗️& Time Series Foundation Model📈& Causality💡 Ex: @PKU1898; @Adobe, UCB, MSRA, Alibaba , Baidu
99K Followers 8K FollowingCompiling in real-time, the race towards AGI.
The Largest Show on X for AI.
🗞️ Get my daily AI analysis newsletter to your email 👉 https://t.co/6LBxO8215l
19K Followers 1K FollowingAgents @Meta MSL TBD Lab. previously posttraining research @OpenAI train LLMs to do things: deep research, chatgpt agent, etc. CS PhD @LTIatCMU
24K Followers 689 FollowingProfessor and Head of Machine Learning Department at @CarnegieMellon. Board member @OpenAI and @Qualcomm. Chief Technical Advisor @GraySwanAI.
44 Followers 21 FollowingI am a postdoctoral scholar at UCLA. Now, I am leading a group for lifelong learning from web data to build a visual knowledge base.
406 Followers 2K FollowingPh.D Candidate @sjtu1896, Intern @Alibaba_Qwen. Exploring Data-Centric AI on LLMs, MLLMs, including data synthesis/pruning/distillation/attribution.
25K Followers 100 FollowingDirector, @PrincetonPLI and Professor @PrincetonCS. Seeks math/conceptual understanding of deep learning and large AI models.
Also on the "other" social network
21K Followers 470 Followingphysics of language models @ Meta (FAIR, not GenAI, not TBD)
🎓:Tsinghua Physics — MIT CSAIL — Princeton/IAS
🏅:IOI x 2 — ACM-ICPC — USACO — Codejam — math MCM
315 Followers 504 FollowingPhd student of @USC' CS. Working with Prof. @yanliu_usc. LLM 🏗️& Time Series Foundation Model📈& Causality💡 Ex: @PKU1898; @Adobe, UCB, MSRA, Alibaba , Baidu
2K Followers 1K FollowingAssistant professor @UMichCSE @UMich; previously @SimonsInstitute @UCBerkeley @Princeton @Tsinghua_Uni. Theoretical and scientific foundations of deep learning.
26K Followers 884 FollowingResearch Scientist Director in Meta FAIR. Reasoning, Optimization and Understanding LLM. Novelist in spare time. PhD in @CMU_Robotics.
38K Followers 565 FollowingAssistant professor at Stanford; Co-founder of Voyage AI (https://t.co/wpIITHLgF0) ;
Working on ML, DL, RL, LLMs, and their theory.
1K Followers 472 Followingassistant professor @JHUCompSci & @JHUCogSci | director of SCAI lab | working on machine social intelligence, embodied AI, and computational social cognition
3K Followers 273 FollowingBoeing Endowed Professor in the Allen School of Computer Science at the University of Washington; @uwcse https://t.co/TSBy7bHgCS
4.4M Followers 3 FollowingOpenAI’s mission is to ensure that artificial general intelligence benefits all of humanity. We’re hiring: https://t.co/dJGr6Lg202
No recent Favorites. New Favorites will appear here.