Blaise Delattre

[MILESDauphine @DauphineMiles]

🍾🍾🍾 Best paper award for @florianflb at #eacl2024 ! For his first paper, this is a good start 😃. A nice collaboration with @mlia_isir .

[MILESDauphine @DauphineMiles]

Congratulations to @florianflb and @AlexVerine for their great new paper accepted at #acl2024 on exploring precision and recall for LLMs: https://t.co/fJIbX10CJw

[MILESDauphine @DauphineMiles]

Bravo @bldelattre for your wonderful poster presented at #iclr. Check the full paper: https://t.co/2rGLPMwh6M

[Chawin Sitawarin @csitawarin]

🗡️ We propose PAL, a new practical black-box jailbreak attack on LLM APIs.

It is a token-level, optimization-based attack like GCG (Zou et al., 2023) but uses only ~3k queries on average to jailbreak GPT-3.5-Turbo! It also succeeds ~50% against Llama-2-7b-chat-hf in 25k queries.