Two Birds, One Projection: Harmonizing Safety and Utility in LVLMs via Inference-time Feature Projection

Existing jailbreak defence frameworks for Large Vision-Language Models often suffer from a safety utility tradeoff, where strengthening safety inadvertently degrades performance on general visual-grounded reasoning tasks. In this work, we investigate whether safety and utility are inherently antagonistic objectives. We focus on a modality induced bias direction consistently observed across datasets, which arises from suboptimal coupling between the Large Language Model backbone and visual encoders. We further demonstrate that this direction undermines performance on both tasks. Leveraging this insight, we propose Two Birds, One Projection, an efficient inference time jailbreak defence that projects cross-modal features onto the null space of the identified bias direction to remove the corresponding components. Requiring only a single forward pass, our method effectively breaks the conventional tradeoff, simultaneously improving both safety and utility across diverse benchmarks.

DiSPA: Differential Substructure-Pathway Attention for Drug Response Prediction

Accurate prediction of drug response in precision medicine requires models that capture how specific chemical substructures interact with cellular pathway states. However, most existing deep learning approaches treat chemical and transcriptomic modalities independently or combine them only at late stages, limiting their ability to model fine-grained, context-dependent mechanisms of drug action. In addition, standard attention mechanisms are often sensitive to noise and sparsity in high-dimensional biological networks, hindering both generalization and interpretability. We present DiSPA, a representation learning framework that explicitly disentangles structure-driven and context-driven mechanisms of drug response through bidirectional conditioning between chemical substructures and pathway-level gene expression. DiSPA introduces a differential cross-attention module that suppresses spurious pathway-substructure associations while amplifying contextually relevant interactions. Across multiple evaluation settings on the GDSC benchmark, DiSPA achieves state-of-the-art performance, with particularly strong improvements in the disjoint-set setting, which assesses generalization to unseen drug-cell combinations. Beyond predictive accuracy, DiSPA yields mechanistically informative representations: learned attention patterns recover known pharmacophores, distinguish structure-driven from context-dependent compounds, and exhibit coherent organization across biological pathways. Furthermore, we demonstrate that DiSPA trained solely on bulk RNA-seq data enables zero-shot transfer to spatial transcriptomics, revealing region-specific drug sensitivity patterns without retraining. Together, these results establish DiSPA as a robust and interpretable framework for integrative pharmacogenomic modeling, enabling principled analysis of drug response mechanisms beyond post hoc interpretation.

Energy-based Preference Optimization for Test-time Adaptation

Test-Time Adaptation (TTA) enhances model robustness by enabling adaptation to target distributions that differ from training distributions, improving real-world generalizability. Existing TTA approaches focus on adjusting the conditional distribution; however these methods often depend on uncertain predictions in the absence of label information, leading to unreliable performance. Energy-based frameworks suggest a promising alternative to address distribution shifts without relying on uncertain predictions, instead computing the marginal distribution of target data. However, they involve the critical challenge of requiring extensive SGLD sampling, which is impractical for test-time scenarios requiring immediate adaptation. In this work, we propose Energy-based Preference Optimization for Test-time Adaptation (EPOTTA), which is based on a sampling free strategy. We first parameterize the target model using a pretrained model and residual energy function, enabling marginal likelihood maximization of target data without sampling. Building on the observation that the parameterization is mathematically equivalent to DPO objective, we then directly adapt the model to a target distribution without explicitly training the residual. Our experiments verify that EPOTTA is well-calibrated and performant while achieving computational efficiency.