Vision language models (VLMs) demonstrate strong zero-shot performance, but often perpetuate social stereotypes in person-centric queries, yielding skewed demographic distributions.Current debiasing methods apply uniform bias corrections across all input queries regardless of their bias sensitivity, creating a fundamental fairness-utility trade-off.Strong debiasing distorts semantically meaningful information in bias-insensitive queries, while weak debiasing fails to mitigate stereotypes in biassensitive ones.This one-size-fits-all approach hampers simultaneously achieving high utility on bias-insensitive queries and fairness on bias-sensitive queries.We introduce Reward-Gated Test-Time Adaptation (RG-TTA), a reinforcement learning-based test-time adaptation framework that selectively applies debiasing based on input sensitivity.RG-TTA adaptively triggers fairness regularization based on the bias sensitivity of each input during testtime policy adaptation, while focusing exclusively on optimizing cross-modal alignment for bias-insensitive inputs.Experiments on fairness benchmarks (e.g., FairFace, UTKFace) demonstrate substantial bias reduction while simultaneously improving zero-shot utility, resolving the trade-off of uniform debiasing.

Hypervisors are increasingly utilized in modern computer systems, ranging from PCs to web servers and data centers. Aside from server applications, hypervisors are also becoming a popular target for implementing many security systems, since they provide a small and easy-to-secure trusted computing base. This paper presents a novel way of using hypervisors to protect application data privacy even when the underlying operating system is not trustable. Each page in virtual address space is rendered to user applications according to the security context the application is running in. The hypervisor encrypts and decrypts each memory page requested depending on the application's access permission to the page. The main result of this system is the complete removal of the operating system from the trust base for user applications' data privacy. To reduce the runtime overhead of the system, two optimization techniques are employed. We use page-frame replication to reduce the number ofcryptographic operations by keeping decrypted versions of a page frame. We also employ lazy synchronization to minimize overhead due to an update to one of the replicated page frame. Our system is implemented and evaluated by modifying the Xen hypervisor, showing that it increases the application execution time only by 3% for CPU and memory-intensive workloads.