Ruidi Chang (@RuidiChang) profile picture

PhD student @RiceCompSci

Joined February 2024
Photo by Ruidi Chang (@RuidiChang)
Ruidi Chang retweeted
⚠️ Fine-tuning LLMs can easily compromise their safety alignment, even with seemingly harmless data! 🚀 We propose ✨GR-SAP✨ to preserve safety alignment with model-synthesized data during fine-tuning. Check our new preprint! 📄 arxiv.org/abs/2603.10243
1
5
2
10
2,157
Ruidi Chang retweeted
📢I'm recruiting a postdoc at Rice University! Looking for someone with a PhD in CS (or a related field) and expertise in LLMs and Trustworthy & Responsible AI. I'll be at #COLM2025 from 10/8-10/10. Happy to chat in person! Feel free to email or DM me if you're interested.😀
3
56
1
247
26,691
Our group is recruiting a postdoc!
📢I'm recruiting a postdoc at Rice University! Looking for someone with a PhD in CS (or a related field) and expertise in LLMs and Trustworthy & Responsible AI. I'll be at #COLM2025 from 10/8-10/10. Happy to chat in person! Feel free to email or DM me if you're interested.😀
3
490
We’re thrilled to present our latest work at ICML 2025! Distribution-wise Intervention is a simple yet powerful technique that learns intervention regions. Come talk to us at ICML if you’re interested—we’d love to chat! #ICML2025
[#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 reparameterization. Given a deterministic intervention function (mean network), we equip it with an independent network (logvar network) to learn its surrounding region. 🎄 Key Takeaway: - D-Intervention is effective at early layers, achieving a drop-in gain compared to its deterministic counterparts at 2% to 4% improvement. - The downstream performance is pretty correlated w/ its learned variance, the more it can explore its surrounding, the more it could achieve better results. - Controlling test-time stochasticity to fulfill the need for creativity in different tasks! Very fun to play with 😀
4
2
10
1,190
🚀 Thrilled to announce SAFR is here! #NAACL2025 Superposition is powerful — but it buries interpretability. We control that! 🧠 Neurons often mix too many features (superposition) — making models a black box. 🎯 SAFR strategically redistributes neurons: · 🧩 Monosemantic for important tokens · 🔗 Polysemantic where interactions matter 🔥 SAFR controls how neurons represent features — without sacrificing accuracy. 💬 Excited to present this work at #NAACL2025 Findings — come chat with us!
1
5
2
39
7,609
We will presenting our paper today! 📍Hall 3, Session B: Oral/Poster 1, Wednesday 4/30/2025, 11:00-12:30 Come and say hi~👋
1
3
234
I'm at #NAACL2025 this week (4/29-5/3), presenting three papers with my students and collaborators.✨ Excited to meet old and new friends, and chat about interpretability, explanation, reasoning, evaluation, and more!☕️
1
7
1
66
5,829