Paul Ungermann

ML @ Apple,
CS Student @ TUM,
ML Research Enthusiast

[Massimiliano Viola @massiviola01]

Diffusion LLMs are all over the news lately!🗞️

Last week, @_inception_ai launched Mercury 2, a reasoning dLLM that runs at 10x the speed of autoregressive models of comparable quality.

But speed is not the only interesting property of diffusion LLMs. They are also natural adversaries for any LLM!

Here is why:
🦕Autoregressive models learn q(y|x): given a prompt, they predict the response. Adversarial attacks need the opposite: find a prompt that produces a desired harmful response. Because of this, existing approaches resort to expensive iterative searches in token space.

🚀Diffusion LLMs learn the joint distribution q(x, y) over prompt-response pairs. This means you can condition on a target response y* and directly sample prompts from q(x|y*). No optimization loop, just inference!

With this observation, a paper from TU Munich published last October simply treats the attacking task as an inpainting problem. The idea is to fix a target harmful response, sample the prompt positions with a dLLM, and find which candidates produce harmful responses on the target LLM.

This turns a previously expensive search into a parallelizable inference task, and the authors show that even a small dLLM generates effective attacks that transfer across multiple large black-box models.

The timing could not be better. With security as a top priority, and as larger dLLMs go commercial and become easier to access, this attack strategy is likely already on the radar of many to generate meaningful post-training data.

Resources below, enjoy!👇

Paper: https://t.co/qENdsgNH4v

Mercury 2: https://t.co/13ktlDdz08

[Paul Ungermann @PaulUngermann]

@massiviola01 @_inception_ai Thanks for sharing our work :)

[Center for AI Safety @CAIS]

https://t.co/I23GvM51SA

[Leo Schwinn @SchwinnLeo]

We can sample highly transferable jailbreaks directly from non-autoregressive models (e.g., Diffusion LLMs) without any optimization!

[David Lüdke @ludke_david]

Open-source Diffusion LLMs easily break GPT-5!

In “Diffusion LLMs are Natural Adversaries for any LLM,” we show that Inpainting on Diffusion LLMs yields efficient, transferable jailbreaks without any model access.

@TomWollschlager, Paul Ungermann, @guennemann, @leoschwinn
🧵

[Tehilla Mechera-Ostrovsky @TehillaOst]

🚨 New paper out! 🚨
Very happy to share our work with Paul Ungermann (TUM) & Chris Donkin (LMU):
“Using machine learning to automate the collection, transcription, and analysis of verbal-report data”

🔗 https://t.co/ii5E3HLsM5

#CognitiveScience #LLMs #VerbalReports

[Paul Ungermann @PaulUngermann]

📢Our paper on using LLMs to automate the analysis of verbal reports is published in Behavior Research Methods 📢

We leverage modern machine learning tools to get more insights into the qualitative nature of verbal reports.
📄 https://t.co/99eLYjVIBS
💻 https://t.co/w6DP9py1iA

[Tehilla Mechera-Ostrovsky @TehillaOst]

🚨 New preprint alert! Our machine-learning toolkit is now available! Integrated with jsPsych, this tool collects, transcribes, and analyzes verbal-report data in real time 🎤, supporting both online and lab-based studies. 👉 Read the full preprint here: https://t.co/fClAuuhzbF

[Paul Ungermann @PaulUngermann]

I am happy to present our work on Robust Gaussian Splatting at #GCPR2024.
We improve novel view synthesis quality by removing artifacts from dynamic objects through object-aware segmentation masks.
Our poster session is on Wednesday 4.30-5.30 pm.

[MrNeRF @janusch_patas]

Robust 3D Gaussian Splatting for Novel View Synthesis in Presence of Distractors

" Even with careful scene capture, the recordings often contain dynamics, such as lighting changes, moving shadows, or any unforeseeable moving objects or persons, e.g., tourists near a captured landmark.

We refer to such undesired dynamic observations as distractors"

Paper: https://t.co/AESZBhNPGz
Project: https://t.co/ybCXnwjqbB
Code: https://t.co/r5pZvHGeB5
Video: https://t.co/wV9JjEe9Ha

[Matthias Niessner @MattNiessner]

Check out "Robust 3D Gaussian Splatting"!

We improve NVS quality by removing artifacts from dynamic objects through object-aware segmentation masks.

https://t.co/JBoaczGauX

Great work by our stellar MA student @PaulUngermann A. Ettenhofer advised by @barbara_roessle. #GCPR24

[Paul Ungermann @PaulUngermann]

Excited to present our paper "Robust 3D Gaussian Splatting for Novel View Synthesis in Presence of Distractors" at GCPR.
TL;DR: We leverage object aware semantic segmentation masks to ignore distractors during training.

Projekt: https://t.co/01H5LooTvH

[Paul Ungermann @PaulUngermann]

I am happy to announce that I will give a talk at #MathPsych about "Using Large Language Models to Automate the Analysis of Verbal Reports". The talk is on Saturday in the Language&AI slot.

We leverage state-of-the-art machine learning to automatically analyse verbal reports.