Steering vectors are a lightweight tool for controlling LLM behavior. However, emerging evidence shows that steering vectors can unintentionally compromise a model's safety mechanisms and increase compliance with harmful requests, while no effective mitigation yet exists. In this work, we show that this safety degradation arises from a separable component in the vector that disrupts the model's safety mechanisms but contributes little to the steering objective. We identify and remove this safety-degrading component, formulating the task as a constrained optimization problem solved through primal-dual updates, subject to preserving the intended steering effect and bounding false refusal. The resulting solution is both interpretable and surgical: the optimization recovers a single direction whose ablation from the steering vector restores model safety with minimal utility cost. Across models, steering behaviors, and attack suites, including unseen attacks types, our method substantially reduces steering-induced safety degradation while preserving the original steering effect with minimal impact on false refusal. Our method offers a post-hoc correction to steering vectors that mitigates their safety cost, and more broadly, it provides a general recipe for applying activation-level model interventions without paying a safety tax.

Activation steering has emerged as a powerful tool to shape LLM behavior without the need for weight updates. While its inherent brittleness and unreliability are well-documented, its safety implications remain underexplored. In this work, we present a systematic safety audit of steering vectors obtained with Contrastive Activation Addition (CAA), a widely used steering approach, under a unified evaluation protocol. Using JailbreakBench as benchmark, we show that steering vectors consistently influence the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks. Across LLM families and sizes, steering the model in specific directions can drastically increase (up to 57%) or decrease (up to 50%) its attack success rate (ASR), depending on the targeted behavior. We attribute this phenomenon to the overlap between the steering vectors and the latent directions of refusal behavior. Thus, we offer a traceable explanation for this discovery. Together, our findings reveal a previously unobserved safety gap in LLMs, marking a trade-off between controllability and safety. We release our code at https://anonymous.4open.science/r/steering-safety.

LLMs are now an integral part of information retrieval. As such, their role as question answering chatbots raises significant concerns due to their shown vulnerability to adversarial man-in-the-middle (MitM) attacks. Here, we propose the first principled attack evaluation on LLM factual memory under prompt injection via χmera, our novel, theory-grounded MitM framework. By perturbing the input given to "victim" LLMs in three closed-book and fact-based QA settings, we undermine the correctness of the responses and assess the uncertainty of their generation process. Surprisingly, trivial instruction-based attacks report the highest success rate (up to 85.3% while simultaneously having a high uncertainty for incorrectly answered questions. To provide a simple defense mechanism against χmera, we train Random Forest classifiers on the response uncertainty levels to distinguish between attacked and unattacked queries (average AUC of up to 96%). We believe that signaling users to be cautious about the answers they receive from black-box and potentially corrupt LLMs is a first checkpoint toward user cyberspace safety.

We propose a modular framework for single-view indoor scene 3D reconstruction, where several core modules are powered by diffusion techniques. Traditional approaches for this task often struggle with the complex instance shapes and occlusions inherent in indoor environments. They frequently overshoot by attempting to predict 3D shapes directly from incomplete 2D images, which results in limited reconstruction quality. We aim to overcome this limitation by splitting the process into two steps: first, we employ diffusion-based techniques to predict the complete views of the room background and occluded indoor instances, then transform them into 3D. Our modular framework makes contributions to this field through the following components: an amodal completion module for restoring the full view of occluded instances, an inpainting model specifically trained to predict room layouts, a hybrid depth estimation technique that balances overall geometric accuracy with fine detail expressiveness, and a view-space alignment method that exploits both 2D and 3D cues to ensure precise placement of instances within the scene. This approach effectively reconstructs both foreground instances and the room background from a single image. Extensive experiments on the 3D-Front dataset demonstrate that our method outperforms current state-of-the-art (SOTA) approaches in terms of both visual quality and reconstruction accuracy. The framework holds promising potential for applications in interior design, real estate, and augmented reality.

In this work, we propose a siamese network based approach to quantify the visual similarity of Japanese Kokatsuji (Old Movable Type) printing blocks, empowering humanities scholars to conduct quantitative analyses of Kokatsuji books.