Niloofar Mireshghallah finds downstream finetuning reactivates verbatim recall of memorized copyrighted books in LLMs, bypassing RLHF

Excited to attend #ICML2026 🇰🇷 soon!! I'll be giving talks, participating on panels and debating @lschmidt3 on future of research and humans post AGI!!

You can find me at @WiMLworkshop @farairesearch @taig_icml, among other places (see👇) papers + links in the thread.

Wed Jul 8 — Panelist at the WiML Symposium, "Career Transitions in ML." Come chat about non-linear paths in ML! http://wiml.org/events/wiml-symposium-@-icml-2026

Sat Jul 11 — Co-organizing MemFM, the 2nd Workshop on the Impact of Memorization on Trustworthy Foundation Models (Room E5–E6).

I'll also give an oral on "Alignment Whack-a-Mole" — finetuning reactivates verbatim recall of copyrighted books. Workshop: http://icml2026memfm.github.io Paper: http://arxiv.org/abs/2603.20957

Missing @icmlconf but if you are in Seoul and attending MemFM attend Niloofar’s talk :-)

@niloofar_mire @lschmidt3 @WiMLworkshop @farairesearch @taig_icml Excited to see you at ICML!!

Fri Jul 10 — Poster: "SMDD-Bench: Can LLMs Solve Real-World Small Molecule Drug Design Tasks?" @ GenBio workshop, Hall D2. w/ Kevin Han, Renfei Zhang, Kathy Wei, Hamed Mahdavi, Amir Barati Farimani Paper: http://arxiv.org/abs/2605.21740

Workshop: http://genbio-workshop.github.io/2026

@niloofar_mire @lschmidt3 @WiMLworkshop @farairesearch @taig_icml See you in ICML!

unfortunately im gonna miss this but if ur there come and attend Niloofar’s talk!