Nils Lukas

A frozen video tokenizer's latent can be factored into controllable and reward-relevant components, unlocking significant downstream planning gains.

Web agents trained with adaptive adversarial simulation achieve 33.6% higher completion rates against unseen prompt injection attacks, even outperforming frontier adversaries.

Generative models are becoming so good at mimicking real images that content alone is no longer a reliable indicator of authenticity, necessitating a shift to reconstruction-based verification.

Diffusion language models can be steered to generate biased outputs by exploiting their iterative denoising process with a closed-loop controller, achieving significantly higher bias injection than fixed steering.

Adversarial attacks can completely neutralize trigger-tag mechanisms designed for misuse detection in open-weight LLMs, rendering them ineffective even with output transformations or weight modifications.

Targeted non-literal suppression in LLMs reveals a critical trade-off between compliance and output quality that could redefine copyright defense strategies.

A shift from artifact matching to evidence-based semantic verification could redefine how we detect AI-generated videos.

Fine-tuning on benign data can paradoxically revive harmful behaviors, revealing a hidden gravitational pull from earlier training phases.