Gert Lek, Abele Malan, Chaoyi Zhu, Pin-Yu Chen, Robert Birke, Lydia Y. Chen

Moves guard models from discriminative to generative classification to create a more robust training objective: rather than predicting a verdict token, LLaDA-Guard asks which safety label better explains the text, spreading supervision over every moderated token. It fine-tunes LLaDA-8B-Instruct with a class-conditional reconstruction objective, leading on average rank across seven held-out safety benchmarks with better calibration.

Camera-ready version and full code coming soon.

Gert Lek, Zixuan Xia, Pin-Yu Chen, Lydia Y. Chen

Shows that as Natural Language Autoencoder explanations of model activations get better at predicting behavior, they increasingly contain unsupported details and writing flaws. Introduces an evaluation framework for information recovery, contextual support, and writing quality, and proposes Flow-NLA, which uses diffusion likelihood bounds to retain the utility gains of point reconstruction while curbing confabulation. Evaluated on Qwen, Gemma, and Apertus.

Gert Lek, Chaoyi Zhu, Pin-Yu Chen, Robert Birke, Lydia Y. Chen

Introduces an Optimal Transport view of activation steering for masked diffusion models, learning a lightweight affine map that transports activation distributions between behaviors. Unifies common steering rules as special cases and improves instruction-following accuracy across LLaDA-Instruct, LLaDA 1.5, and Dream-Instruct.

Gert Lek, Nicolas van Schaik, Chaoyi Zhu, Pin-Yu Chen, Robert Birke, Lydia Y. Chen

Extends SAM2 for image forgery localization with perturbation-driven forensic embeddings, lightweight feature adapters, and a learnable prompt module. Introduces AutoEditForge, an automated diffusion edit generation pipeline. Achieves SOTA on common benchmarks and modern editing models such as NanoBanana and Qwen-Image-Edit.

Gert Lek, Chaoyi Zhu, Pin-Yu Chen, Robert Birke, Lydia Y. Chen

Extends SAM with blur-driven forensic embedding signals, hierarchical learnable prompts, and lightweight adapters for automatic forgery mask generation. Outperforms prior methods on MagicBrush and CoCoGlide.