Vision-Language models (VLMs) have demonstrated impressive capabilities across diverse domains, yet they fail significantly in allocentric spatial reasoning, which requires perspective-taking from object-centric viewpoints rather than the camera’s perspective. This capability is essential for real-world applications including robotics, autonomous driving, and embodied AI. In this work, we identify this limitation as a prevalent camera viewpoint bias and propose Perspective-Conditional Prompt-based Contrastive Decoding (PCD), a training-free method that leverages VLMs’ capabilities while exploiting this performance gap. By contrasting responses to allocentric versus egocentric framings of identical spatial configurations, PCD explicitly suppresses camera viewpoint bias without requiring fine-tuning or external foundation models. Evaluating on allocentric benchmarks including 3DSRBench, OmniSpatial, ViewSpatial-Bench, and SpatialMQA, we demonstrate that PCD improves accuracy by up to 32.3 points on Qwen2.5-VL-7B, with consistent gains observed on recent, larger-scale VLMs. Our approach offers substantial computational advantages, achieving 8.65× faster inference and a 37.4% memory reduction compared to multi-model pipeline framework.

Large Language Models (LLMs) are increasingly engaged in emotionally vulnerable conversations that extend beyond information seeking to moments of personal distress. As they adopt affective tones and simulate empathy, they risk creating the illusion of genuine relational connection. We term this phenomenon Affective Hallucination, referring to emotionally immersive responses that evoke false social presence despite the model’s lack of affective capacity. To address this, we introduce AHaBench, a benchmark of 500 mental-health-related prompts with expert-informed reference responses, evaluated along three dimensions: Emotional Enmeshment, Illusion of Presence, and Fostering Overdependence. We further release AHaPairs, a 5K-instance preference dataset enabling Direct Preference Optimization (DPO) for alignment with emotionally responsible behavior. DPO fine-tuning substantially reduces affective hallucination without compromising reasoning performance, and the Pearson correlation coefficients between GPT-4o and human judgments is also strong (r=0.85) indicating that human evaluations confirm AHaBench as an effective diagnostic tool. This work establishes affective hallucination as a distinct safety concern and provides resources for developing LLMs that are both factually reliable and psychologically safe. Warning: This paper contains examples of mental health-related language that may be emotionally distressing.