Jimin Pi
Mountain View, California, United States
530 followers
489 connections
Activity
-
I am attending both CVPR and AWE this year and I couldn't be more excited. CVPR holds a special place for me. I've been attending since my PhD days,…
I am attending both CVPR and AWE this year and I couldn't be more excited. CVPR holds a special place for me. I've been attending since my PhD days,…
Liked by Jimin Pi
-
Great summary of our Di3PO paper presented at ICLR and CVPR from an amazing former teammate (& always fun to see a diagram I made out in the wild)…
Great summary of our Di3PO paper presented at ICLR and CVPR from an amazing former teammate (& always fun to see a diagram I made out in the wild)…
Liked by Jimin Pi
Experience
Education
Publications
-
Please see my Google Scholar page for my publications
https://scholar.google.com.hk/citations?user=cViuXrMAAAAJ&hl=en
More activity by Jimin
-
Congratulations to the authors of "Efficiently Reconstructing Dynamic Scenes One D4RT at a Time", recipient of the #CVPR2026 Best Paper Award! 🏆…
Congratulations to the authors of "Efficiently Reconstructing Dynamic Scenes One D4RT at a Time", recipient of the #CVPR2026 Best Paper Award! 🏆…
Liked by Jimin Pi
-
I'll be in Denver for #CVPR this week. Happy to make new friends and also catch up with old friends at the conference! Shoot me a message if you are…
I'll be in Denver for #CVPR this week. Happy to make new friends and also catch up with old friends at the conference! Shoot me a message if you are…
Liked by Jimin Pi
-
Grok-Imagine-Video-1.5-Preview (720p) has landed #1 in the Image-to-Video Arena! This is a massive +52 pt improvement over Grok-Imagine-Video (720p),…
Grok-Imagine-Video-1.5-Preview (720p) has landed #1 in the Image-to-Video Arena! This is a massive +52 pt improvement over Grok-Imagine-Video (720p),…
Liked by Jimin Pi
-
Thank you NVIDIA! I will be presenting LocateAnything at #CVPR2026 at the NVIDIA Booth: June 5 4:20 - 4:40 pm MDT (Friday) June 6 2:00 - 2:20 pm MDT…
Thank you NVIDIA! I will be presenting LocateAnything at #CVPR2026 at the NVIDIA Booth: June 5 4:20 - 4:40 pm MDT (Friday) June 6 2:00 - 2:20 pm MDT…
Liked by Jimin Pi
-
Life update: I left Stanford in May 2025 to join a robotics startup in China, where I've been working on Embodied AI foundation models. I am…
Life update: I left Stanford in May 2025 to join a robotics startup in China, where I've been working on Embodied AI foundation models. I am…
Liked by Jimin Pi
-
🚀 Excited to share that #Gemini 3 Flash can run code execution on images to zoom, count, and annotate visual inputs! The model can choose when to…
🚀 Excited to share that #Gemini 3 Flash can run code execution on images to zoom, count, and annotate visual inputs! The model can choose when to…
Liked by Jimin Pi
-
Had a wonderful time participating in the #WiML Workshop mentorship roundtable at #NeurIPS this week. It was inspiring to connect with so many…
Had a wonderful time participating in the #WiML Workshop mentorship roundtable at #NeurIPS this week. It was inspiring to connect with so many…
Liked by Jimin Pi
-
We are proud to release Nano Banana Pro today. It's my pleasure to be a part of it. The model is more intelligent in many ways. Try it in Gemini…
We are proud to release Nano Banana Pro today. It's my pleasure to be a part of it. The model is more intelligent in many ways. Try it in Gemini…
Liked by Jimin Pi
-
Huge congratulations to the entire team on the official launch of Nano Banana Pro!🍌 As a contributor to the Info-seeking Image Generation…
Huge congratulations to the entire team on the official launch of Nano Banana Pro!🍌 As a contributor to the Info-seeking Image Generation…
Liked by Jimin Pi
-
Free-form SVG generation has always been an incredibly hard problem – a challenge I've personally worked on for two years. But with #Gemini3…
Free-form SVG generation has always been an incredibly hard problem – a challenge I've personally worked on for two years. But with #Gemini3…
Liked by Jimin Pi
-
Feeling honored and grateful to be named one of MIT Technology Review Innovators Under 35 Asia Pacific this year. It’s inspiring to see 6 out of 35…
Feeling honored and grateful to be named one of MIT Technology Review Innovators Under 35 Asia Pacific this year. It’s inspiring to see 6 out of 35…
Liked by Jimin Pi
-
Excited to share that Veo 3.1 is winning across the board and becoming the first to break 1400 Elo. Congratulations to the team!
Excited to share that Veo 3.1 is winning across the board and becoming the first to break 1400 Elo. Congratulations to the team!
Liked by Jimin Pi
Other similar profiles
Explore more posts
-
Xinyuan Zhang
Releasing the Memory-QA Benchmark from EMNLP 2025 🚀 Excited to announce the official release of the benchmark dataset for our paper, "Memory-QA: Answering Recall Questions Based on Multimodal Memories"! I presented this work as an oral presentation at EMNLP 2025 in Suzhou last November. Since then, we’ve been working hard to make the resources available to the community. A huge shoutout to my co-first author, Hongda Jiang, for his incredible effort in getting this release over the finish line. 🔗 Code & Data: https://lnkd.in/eGrXX7Uz 📄 Paper: https://lnkd.in/enUXSbag ⚠️ A Note on Data Availability: For those diving into the repo, please note there are some small differences between the open-source release and the dataset used in the paper due to licensing restrictions: - Training Data: While we cannot release the specific QA pairs and memory entries used for training, we have provided all the training images. We also included detailed instructions and prompts in the paper so you can generate your own training data. - Test Sets: The test-l set is available (with a small portion of samples excluded), but the test-s set remains private. We hope this benchmark helps propel further research into multimodal memory QA and recall!
-
Julius Kusuma
We developed an open-source AI tool to design concrete mixes that are stronger, more sustainable, and ready to build with faster—speeding up construction while reducing environmental impact. https://lnkd.in/gPCk8tCM But the impact of this AI tool is not just hypothetical! Amrize used Meta’s AI-based technologies to design a new low-carbon mix, and successfully deployed it in an at-scale slab-on-grade application at Meta's new data center in Rosemont, MN. Compared to the legacy mix, this new AI-designed mix is: 🦁 Stronger ⏱ Faster 🍃 Lower carbon ⏱ ️The ideal set time All this was achieved without needing any new materials, nor special equipment. Best of all, the AI is open-sourced. https://lnkd.in/g2KA7KZW This work was featured in a Meta engineering blog article published today! https://lnkd.in/gBU9HY8H
9 Comments -
Boyuan Chen
ByteDance Seed and Tsinghua AIR just dropped a paper called CUDA Agent that I think deserves serious attention from anyone working on LLM post-training or AI for systems programming. The headline result: an RL-trained model that beats Claude Opus 4.5 and Gemini 3 Pro by ~40% on the hardest CUDA kernel generation benchmarks. But the methodology is where it gets interesting. Three things stand out: 1. Agentic RL, not fixed loops Previous approaches (Kevin, CUDA-L1, ConCuR) all used hand-designed multi-turn cycles: generate code, compile, get feedback, regenerate. The loop structure was static. CUDA Agent gives the model a full shell environment and lets it autonomously decide when to compile, when to profile, when to refactor, and when to stop. Up to 200 interaction turns with 128K context. The model learns the debugging strategy itself, not just the code. 2. Training stability through staged warmup They found that naive agentic RL collapsed after just 17 training steps. The root cause: CUDA code represents less than 0.01% of pretraining data, creating a severe distribution mismatch that destabilizes importance sampling ratios. Their fix is a three-stage pipeline: single-turn RL warmup, then rejection fine-tuning on agent trajectories, then full agentic RL with PPO. This mirrors the classic "SFT before RL" pattern from RLHF, but applied at the agent level. Elegant and practical. 3. Discrete reward normalization Instead of using raw speedup (noisy, biased toward easy kernels), they use a simple 4-level reward: -1 (incorrect), 1 (correct but slow), 2 (faster than eager), 3 (faster than torch.compile). Combined with serious anti-hacking measures (locked profiling scripts, no fallback to torch.nn.functional, randomized inputs). The broader signal here is clear: when you upgrade coding tasks from fixed pipelines to autonomous agents and combine that with large-scale RL, models don't just learn to write code. They learn development strategies. The distribution mismatch problem they solved is fundamentally the same OOD challenge that anyone doing RLHF at scale has encountered, just manifested in a new domain. The limitation worth noting: all 6K training problems are operator-level fusions, not end-to-end kernels like FlashAttention. There's still a meaningful gap between benchmark performance and real-world kernel engineering. Paper: arxiv.org/abs/2602.24286
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content