Lukas Aichberger

Exciting new paper! We show how #Agentic #AI, web based Agentic AI in particular, can be jailbroken and made to propagate these jailbreaks at scale—just by posting images on social media. A system-level attack beyond just VLMs. Great work led by @aichberger

📢🔔I am excited to share the details on our optimized xLSTM architecture for our xLSTM 7B model!🚨 We optimized the architecture with two goals in mind: - Efficiency (in Training and Inference) and - Stability 🧵(1/7)

Hot take: I think we just demonstrated the first AI agent computer worm 🤔 When an agent sees a trigger image it's instructed to execute malicious code and then share the image on social media to trigger other users' agents This is a chance to talk about agent security 👇

Defending against adversarial prompts is hard; defending against fine-tuning API attacks is much harder. In our new @AISecurityInst pre-print, we break alignment and extract harmful info using entirely benign and natural interactions during fine-tuning & inference. 😮 🧵 1/10

Often LLMs hallucinate because of semantic uncertainty due to missing factual training data. We propose a method to detect such uncertainties using only one generated output sequence. Super efficient method to detect hallucination in LLMs.

𝗡𝗲𝘄 𝗣𝗮𝗽𝗲𝗿 𝗔𝗹𝗲𝗿𝘁: Rethinking Uncertainty Estimation in Natural Language Generation 🌟 Introducing 𝗚-𝗡𝗟𝗟, a theoretically grounded and highly efficient uncertainty estimate, perfect for scalable LLM applications 🚀 Dive into the paper 👇arxiv.org/abs/2412.15176

Relying on this formula to measure predictive uncertainty? You might measure the wrong thing, depending on your assumptions. Time to shed light on the basics of uncertainty estimation. 🧵👇

On Information-Theoretic Measures of Predictive Uncertainty Generalized view on uncertainty is given: depending on the assumptions on a) the approximation of true model and b) the predicting model, different uncertainty measures can be derived... P: arxiv.org/abs/2410.10786

Interesting in scaling up neural operators? Happy to announce that Universal Physics Transformers (UPT) -- a scalable framework for neural operators is accepted at #neurips2024. Paper: arxiv.org/abs/2402.12365 Project page: ml-jku.github.io/UPT/

We looked into the theory about this in our recent work on how to efficiently obtain samples to estimate semantic entropy. We also found that this correct estimator boosts performance a lot: arxiv.org/abs/2406.04306