Check out my mentee's latest work on LLM Router attack! First work on this topic to the best of our knowledge. Read our paper at: zhijing-jin.com/files/papers/2… Great job to Aly M. Kassem!🎉 Intelligent Systems U of T Department of Computer Science

Very useful thread — unfortunately, I learned it the hard way.

Finally, the crazy weeks of NeurIPS ddl, ICCV rebuttal and EMNLP ddl have passed. Now it's time to take a rest till Sep!

We observed similar biased behavior when evaluating LLM routers, even with commercial solutions such as Amazon Bedrock By biases, I mean that the router tends to favor certain categories or keywords, consistently directing them to the more powerful model arxiv.org/abs/2504.07113

🎉 Thrilled that our work has been accepted at #EMNLP2025 (Main Conference)! TL;DR: We propose a framework to predict & explain unintended side effects in models (e.g., emergent toxicity, forgotten knowledge) using OOD data. Huge thanks to Golnoosh Farnadi @NeurIPS2024, Negar Rostamzadeh, and Zhuan Shi 🚀