Xiangzhe Xu
@XiangzheX
Followers
115
Following
53
Statuses
30
🔍 Problem: Code LLMs are powerful, but they often exhibit insecure coding practices, posing real-world security risks. How can we align them with secure coding practices without sacrificing performance? 💡 Our Solution: PROSEC – A lightweight, proactive post-training stage security alignment technique that systematically improves models by identifying weaknesses and refining them to follow secure coding practices without requiring extensive retraining. ✨ Key Highlights: ✅ A security-related instruction dataset that exposes 25× more insecure practices from a model, enhancing security learning ✅ 25.2% to 91.4% security improvement without harming code quality ✅ Efficient post-training stage alignment – enhances security without redoing instruction-tuning
2
0
9
RT @natolambert: This is why we use open-source tools that have large communities to implement all 37 of these details and you can stand on…
0
32
0
We released the vulnerability-inducing instruction dataset v0.1 today, at Will release model-specific alignment datasets and the aligned models in following weeks
0
0
1
RT @AymericRoucher: Introducing open-Deep-Research by @huggingface ! 💥 Deep Research from @OpenAI is really good... But it's closed, as us…
0
584
0
RT @dharmesh: MY MIND HAS JUST BEEN BOGGLED. I just tried the early release of OpenAI's new DeepResearch feature (rolling out later today…
0
357
0
RT @BigComProject: Introducing 🏟️SWE Arena: An Open Evaluation Platform for Vibe Coding Unlike the current frontend-dev applications like…
0
28
0
RT @natolambert: Why reasoning models will generalize DeepSeek R1 is just the tip of the ice berg of rapid progress. People underestimate…
0
67
0
RT @Lin0Tan: Nova has been accepted to ICLR 2025! The first author @NanJiang719 is on the academic job market! #BinaryAnalysis #LLM #securi…
0
4
0
RT @dharmesh: I love APIs, but as we move towards a more agent-based world, I think we'll see forward-thinking products and platforms offer…
0
196
0
RT @ZihengJiang: I'm excited to share the key highlights of ByteDance Doubao-1.5-pro! Here are some technical takeaways: - Outperforms dee…
0
16
0
RT @ShiweiFeng3: 🎉 Excited to announce that our paper "ROCAS: Root Cause Analysis of Autonomous Driving Accidents via Cyber-Physical Co-mut…
0
1
0
RT @MingweiZheng2: 🎉Excited to announce our paper: “ParDiff: Practical Static Differential Analysis of Network Protocol Parsers” won #OOPS…
0
7
0
RT @QingkaiS: My co-author and good friend @XiangzheX will present our work in @USENIXSecurity tomorrow. This paper presents a method of in…
0
1
0
RT @ShiweiFeng3: Enjoy the power of self-supervised learning, yet stay safe from attackers. Check out our #CVPR23 paper, 'Backdoor Detectio…
0
2
0
@i2huer Actually took a few minutes to figure out why the ground truth makes sense… I didn’t notice the recursive call at the beginning.
0
0
0