Gautam Kamath

TL;DR: improved training-inference trade-off of drifting models Faster training & comparable FID, costing increased memory usage First author Ali Falahati, co-supervised w/ @elliot-creager.bsky.social & Shubhankar Mohapatra Paper: arxiv.org/abs/2605.12183 Code: github.com/Mort…

DriftXpress: Faster Drifting Models via Projected RKHS Fields arxiv.org

ICML page: https://t.co/FgHGO3rkI0 I'll be on a panel at the KAIST event (https://t.co/vG59MKb34A), and doing a few other things behind the scenes. I'll have @TmlrOrg stickers for AEs, expert reviewers, spotlight & outstanding certifications. Will announce where to get them!

ICML Tutorial Unlearning Data at Scale icml.cc

Lots more in the paper: how does DPO fit into the picture? What if attackers have different goals? etc. Paper: arxiv.org/abs/2606.04929 Code: github.com/jcksanderson... Led by Jack Sanderson (jcksanderson.com), w/ Yihan Wang, Xiaoqian Lu, co-supervised w/ Yiwei Lu

Sequential Data Poisoning in LLM Post-Training arxiv.org

What about poisoning PPO? A remarkable paper of @javirandor.com and @floriantramer.bsky.social (arxiv.org/abs/2311.14455) shows that just 0.5% poison is enough to break a reward model (L)! Again, fear not: somehow, it takes a (high) 5% poisoning before it transfers to the RLHF…

Universal Jailbreak Backdoors from Poisoned Human Feedback arxiv.org

Now on arXiv: arxiv.org/abs/2606.01849

ContinuousBench: Can Differentially Private Synthetic Text Improve Capabilities? arxiv.org

In the last 48h: - Jr researcher asked me wheter to use AI in making talks - Saw two talks, with AI {slop, enhanced} slides Collected my thoughts and wrote a post. Tl;dr: don't steal your own thinking, don't remove *you* from your talks. Also, give a &#@% about your talks.

if you get caught submitting AI slop to arxiv, the punishment should be generational aura loss

I think frontier AI labs should hire people who either: - at least pretend to care about the people affected by their products - can make good jokes? I talk to brilliant young people every day, terrified about the future. This callousness from those inside is sad.

It's so cringe when real people I otherwise know and respect post obvious AI slop on social media, particularly when they're (supposedly) expressing their feelings. Authenticity is so rare and valuable these days, and it's sad to see people just cede it from the get-go

"People submit too many papers to ML conferences" Meanwhile, at IEEE Transactions on Wireless Communications: "You may not submit more than 36 papers per year." Apparently this is a new policy. I would be super curious to see the stats on how many people are submitting more.

When I was a junior grad student, it was a huge honour to be asked to review. I spent a week reading every detail of a paper. Still in theory, you'll get senior faculty giving thoughtful reviews for STOC/FOCS. Sadly, ML requires holding people hostage until they do their reviews😢

We were the runner up for the Caspar Bowden PETS award. This recognized our work on differentially private fine tuning of language models. It was a big effort by a sizeable group across NA, Europe, and Asia. Thanks to the committee for the recognition.

Workshop on Responsibly Enabling Data for Foundation Models at #COLM2026 October 9 in SF "Unlocking sensitive data sources responsibly for the next generation of AI" - Amazing invited speakers 😍 - Submission deadline: June 23 🗓️ - Do *you* want to be a PC member? 👈 @colmweb.org

Taylor Swift really needs to be more thoughtful about scheduling around ML conferences. First her Eras Tour finale clashed with #NeurIPS2024. And now her wedding conflicts with the start of #ICML2026? No regrets about turning down the wedding invite. See you soon in Seoul!

🧵Feeling safe against data poisoning in post-training? Think again! Individual components of LLM post-training pipelines are surprisingly robust to data poisoning attacks. In work led by Jack Sanderson (co-advised w Yiwei Lu), we show they crumble when attacked together. 1/n