Chengzhi Liu
@liuchen02938149
Research：Multimodal language model & Truthworthy AI Incoming CS Phd@UCSB
Multimodal LLMs amplify hallucination during extended reasoning by drifting from visual input. The paper quantifies this issue, finds optimal reasoning length varies, and introduces RH-AUC metric and RH-Bench benchmark for systematic evaluation. Methods 🔧: → Reasoning length…
🛡️ Improved Safe for Large Reasoning Models! 🧠 How can we better align large reasoning models (LRMs) against unseen jailbreaks and harmful prompts? We present SafeKey — a LRM alignment method that helps activate the aha-moment of safety reasoning. 🔥 Key Points: 1️⃣ Aha-moment…
𝘏𝘶𝘮𝘢𝘯𝘴 𝘵𝘩𝘪𝘯𝘬 𝘧𝘭𝘶𝘪𝘥𝘭𝘺—𝘯𝘢𝘷𝘪𝘨𝘢𝘵𝘪𝘯𝘨 𝘢𝘣𝘴𝘵𝘳𝘢𝘤𝘵 𝘤𝘰𝘯𝘤𝘦𝘱𝘵𝘴 𝘦𝘧𝘧𝘰𝘳𝘵𝘭𝘦𝘴𝘴𝘭𝘺, 𝘧𝘳𝘦𝘦 𝘧𝘳𝘰𝘮 𝘳𝘪𝘨𝘪𝘥 𝘭𝘪𝘯𝘨𝘶𝘪𝘴𝘵𝘪𝘤 𝘣𝘰𝘶𝘯𝘥𝘢𝘳𝘪𝘦𝘴. But current reasoning models remain constrained by discrete tokens, limiting their full…
[2] Multimodal Situational Safety. April 25, 3:00-5:30 pm in Hall 3 + Hall 2B #538. Unfortunately, @KaiwenZhou9 cannot attend ICLR due to the visa issue, but welcome to meet my incoming PhD student @liuchen02938149 (the co-first author) and me there! x.com/kaiwenzhou9/st…
Could not attend #ICLR2025 🥲. But @liuchen02938149 will present our Multimodal Situational Safety paper on April 25, 3:00-5:30 pm in Hall 3 + Hall 2B #538. Welcome to check it out!