Feng Luo
@FengLuo895614
Incoming PhD student @ Rice
[#ICML2025] If steering vector can be multiplied by a scalar to control its effect, why it should be a vector instead of a steering "region"? We introduce distribution-wise intervention, a simple method that directly learns the latent intervention region through…
🚀 Can LLMs stop overthinking when detailed reasoning isn't needed? Excited to share our latest work on LLM reasoning: AutoL2S 🧠⚡ 📄 Paper: arxiv.org/abs/2505.22662 🤖 Model: huggingface.co/amandaa/AutoL2… LLMs often overthink—generating unnecessarily long CoTs even for easy…
🚀 Thrilled to announce that our SPORTU (w/@hanjie_chen) has been accepted to #ICLR 2025! 🔢 12,948 total QA pairs: 900 text MCQs with human‑annotated explanations & 12,048 video QAs 📹 Video: 1,701 slow‑motion clips across 7 sports with 300 multi‑camera clips that cause model…
LLM Post-Training: A Deep Dive into Reasoning Large Language Models
New results just dropped 🥳! We have integrated GPT-4.5 and Gemini-2.0-flash in our gaming agents and test them on Super Mario Bros. ⚔️ GPT-4.5 struggles due to high latency, Gemini-2.0-flash performs significantly better than Gemini-1.5-pro, on par with Claude-3.5. Enjoy! 🎮
👍👍👍
GRM has been accepted to #neurips2024 ! 📢 We show that text generation can act as a regularizer for LLM reward models, improving accuracy on OOD tasks and mitigating over-optimization in RLHF at a minimal additional cost. Check out our Arxiv version at arxiv.org/abs/2406.10216
Dall-E 3 now allows you to copy the style of an image. You can generate an unlimited number of images in the same style. Here's how to do it directly in ChatGPT with a new parameter:
3. Create a new image You can now create as many images as you like in the same style. Prompt: "Create an image of [topic] in the style of image [gen_id]" Example: "Create an image of a woman on a bike in the style of image UBz2F4d0oLA5bkrL" Dall-E 3 will understand which…