Basim Azam

Latent-iterative reasoning in VLA models may actually decrease robustness, collapsing under perturbations while simpler reasoning approaches hold strong.

Introspective attention modulation can significantly enhance the safety of T2I models without sacrificing quality, outperforming existing methods like concept erasure.

Injecting intermediate text representations into the denoising process can dramatically enhance the alignment of text-to-image models with challenging prompts for unique objects.

Existing benchmarks fail to capture the retrieval challenges posed by visually homogeneous documents, but Invoice Haystack reveals the stark performance drop of current models in this critical area.

Retraining just the classifier head of a frozen feature extractor can be dramatically improved by meta-learning feature-space augmentations that target hard examples, leading to state-of-the-art robustness against spurious correlations.