🎥 Video is already a tough modality for reasoning. Egocentric video? Even tougher! It is longer, messier, and harder. 💡 How do we tackle these extremely long, information-dense sequences without exhausting GPU memory or hitting API limits? We introduce 👓Ego-R1: A framework

北京时间8月9日上午10点，#青稞Talk 第70期，南洋理工大学MMLab博士生田淑琳 @shulin_tian 将直播分享 《Evaluation Agent：面向视觉生成模型的高效可提示的评估框架》。#青稞社区 #acl2025 青稞社区主页： https://t.co/F50LarR32L ｜ https://t.co/uRGp8rA1SX

Some say evaluation matters more than training in the second half of the AI game. But for generative models—especially diffusion—evaluation is still slow, rigid, and expensive. Enter Evaluation Agent: a human-like, multi-round framework that evaluates with just a few samples.

It's even better! Evaluation Agent is selected as🏆 SAC Highlights Award 🏆 @aclmeeting Huge kudos to my amazing coauthors: @Zhang_Fan_ , @ziqi_huang_ , Yu Qiao, @liuziwei7 Let's make evaluation of GenAI model a more efficient, human-like process 🚀 Code:

🏆 SAC Highlights Award #ACL2025! 🏆 Our work "𝗘𝘃𝗮𝗹𝘂𝗮𝘁𝗶𝗼𝗻 𝗔𝗴𝗲𝗻𝘁: Efficient and Promptable Evaluation Framework for Visual Generative Models" has been selected for the 𝗦𝗔𝗖 𝗛𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀 𝗔𝘄𝗮𝗿𝗱 at ACL 2025! @aclmeeting Code:

🤩Excited to share that our 🤖Evaluation Agent🤖 has been selected as **SAC Highlights Award** @aclmeeting #ACL2025 "Evaluation Agent" employs human-like strategies for efficient, dynamic, multi-round evaluation - Project: https://t.co/fnx7zzTbes - Code:

Honored that Our Evaluation Agent was selected as a SAC Highlight at ACL 2025! 🎉 Huge thanks to the chairs for the recognition! Excited to see growing interest in agent-based evaluation and hope it leads to further progress.

Slides are available here:

Our session will start at CEST 12:00pm (in 30 mins)~

Our paper Evaluation Agent: Efficient and Promptable Evaluation Framework for Visual Generative Models has been awarded an Oral Presentation at #ACL2025! 🎉 Due to some unforeseen reasons, none of the authors could attend onsite in Vienna... But you can still join us virtually:

New on Hugging Face: Ego-R1 Reasoning over ultra-long egocentric videos with Chain-of-Tool-Thought, orchestrated by a RL agent!

5. Ego-R1: Chain-of-Tool-Thought for Ultra-Long Egocentric Video Reasoning 🔑 Keywords: Ego-R1, Chain-of-Tool-Thought, reinforcement learning, ultra-long egocentric videos, video QA benchmark 💡 Category: Multi-Modal Learning 🌟 Research Objective: - To develop Ego-R1, a

🎬Ultra-Long Egocentric Video Reasoning🎬 😎Ego-R1😎 reasons over ultra-long (i.e., in weeks) egocentric videos, which leverages a **Chain-of-Tool-Thought (CoTT)** process trained via reinforcement learning (RL) - Project: https://t.co/M7OWcd7nFS - Code:

Ego-R1: Chain-of-Tool-Thought for Ultra-Long Egocentric Video Reasoning Reasoning over weeks-long video, with a mix of CoT and using tools

My first "Ego" involvement 🤣 Ego-vision, Ego3D/4D are at the heart of affordance discovery, interaction exploration, and embodied AI!

Ego-R1 Chain-of-Tool-Thought for Ultra-Long Egocentric Video Reasoning

[Please help upvote our Huggingface paper 🤍] https://t.co/VEnXqRHKX1

6/🧵 Acknowledgements Kudos to all coauthors and supervisors: @RuiqisNotes, @AllKend7690, @penghaowu2, @dyhTHU, Xiuying Wang, @JingkangY, @richardzhangsfu, Hongyuan Zhu, @liuziwei7 Cheers to the EgoLife team, our supporting labs, and the friends whose thoughtful discussions