🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

🔥NEW ICLR 2025 PAPER! ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models paper: arxiv.org/abs/2504.06838 poster: #377, Poster session 2, Thursday 3-5:30pm, Hall 3 + Hall 2B thread 🧵

The era of generalist AI paves the way to private AI and adaptation. Generalist AI solves IMO problems better than any human. Yet on my email threads it produces rubbish. I cannot fully delegate my work pipeline to AI. This happens because of adaptation issues; we must bring

Glad to be at #ICLR2025! Ankit Sonthalia and I are presenting our paper on star-shapeness of deep learning solution sets today at poster session 5 (poster #132)! Feel free to drop by and talk about the structure of the models' weight space, generalization and uncertainty!

I will be in person at #NAACL2025 🌵🇺🇸 to present Scaling Up Membership Inference: When and How Attacks Succeed on LLMs. Come and say hi 👋 if you want to know how to proof if an LLM was trained on a data point!

here is my take after skimming through the 18 page article- collection-level mia can work even though gpt-style models that see only 4 k–200 k tokens at once. what to need basically is to chunk a ~1 million token corpus, score each chunk (loss, confidence, embedding drift) and

Do you want to prove that your copyrighted document/corpus was trained by an LLM? Come to poster 46 #NAACL2025

I'm so proud! Haritz Puerto

🔴 LIVE NOW! researchtrend.ai/zoom 🤖 LLMAG Session (16:00 - 17:00 UTC) 🖼️ VLM Session (17:00 - 18:00 UTC) 🎯 FedML Session (18:00 - 19:00 UTC)

Congrats!!

I am pleased to announce a new version of my RL tutorial. Major update to the LLM chapter (eg DPO, GRPO, thinking), minor updates to the MARL and MBRL chapters and various sections (eg offline RL, DPG, etc). Enjoy! arxiv.org/abs/2412.05265

🚨 Paper Alert: “RL Finetunes Small Subnetworks in Large Language Models” From DeepSeek V3 Base to DeepSeek R1 Zero, a whopping 86% of parameters were NOT updated during RL training 😮😮 And this isn’t a one-off. The pattern holds across RL algorithms and models. 🧵A Deep Dive

After 6 NeurIPS submissions, 5 ICCV rebuttals, and 1 EMNLP submission, I’m trying to regain sanity. Handling 400+ emails - Done. Now, let's sort out those 169 todos. AI says the estimated average duration for each item is 2.9 hours.

Reasoning models are quite verbose in their thinking process. Is it any good? We find out that it enables reasoning models to be more accurate in telling what they know and don’t know (confidence)! Even non-reasoning models can do it better if they mimic the verbose reasoning! 👀

🔴 Live now ! researchtrend.ai/zoom

Congrats 🎉 Happy to announce a paper with Yujin, Arnas, and Anna.  It was such a fun collaboration. Our paper tackles a fundamental question raised by the Generative AI Paradox (ICLR 2024, arxiv.org/abs/2311.00059): “What it can create, can it also understand?” Some recent

Recently updated my website seongjoonoh.com using community tags from ResearchTrend.AI. It reflects the evolution of my research interests and highlights my current focus. Visitors can browse topics of interest with ease. At ResearchTrend.AI, we're

🔴 LIVE NOW at researchtrend.ai/zoom We've got great speakers: Yuntao Du from Purdue University Shaina Raza, PhD from Vector Institute David Fan from Meta Youssef Allouah from EPFL We cover fresh papers on LLM Agents, VLM, & Federated Learning. Think of it as a giant reading

the new administration in Korea has managed to bring onboard Jung-Woo Ha as the secretary of AI and future technologies to directly report to and advise the President. some real AI researcher, developer and director! a great job!

🔴 LIVE NOW at researchtrend.ai/zoom Event agenda: researchtrend.ai/social-events/… ResearchTrend.AI #LLM #VLM #FederatedLearning #AI