Large Language Models for Simulating Professions (SP-LLMs), particularly as teachers, are pivotal for personalized education. However, ensuring their professional competence and ethical safety remains a major challenge, as existing benchmarks fail to measure role-playing fidelity or address the unique teaching harms inherent in educational scenarios. To address this gap, we propose EduGuardBench, a dual-component benchmark that evaluates professional fidelity through the Role-playing Fidelity Score (RFS) and diagnoses harms specific to the teaching profession. It also probes safety vulnerabilities using persona-based adversarial prompts targeting both general harms and academic misconduct, with metrics such as Attack Success Rate (ASR) and a three-tier Refusal Quality assessment. Extensive experiments on 14 leading models reveal a stark polarization in performance. While reasoning-oriented models generally demonstrate higher fidelity, incompetence remains the dominant failure mode across most models. Adversarial testing uncovered a counterintuitive scaling paradox, where mid-sized models appear more vulnerable, challenging monotonic safety assumptions. Notably, we identify an Educational Transformation Effect, where the safest models convert harmful requests into teachable moments through ideal educational refusals. This ability is strongly negatively correlated with ASR, revealing a new dimension of advanced AI safety. EduGuardBench thus provides a reproducible framework for holistic assessment of professional, ethical, and pedagogical alignment, uncovering dynamics critical to deploying trustworthy AI in education.

Large Vision-Language Models (LVLMs) have shown impressive performances on multimodal tasks, yet remain fundamentally constrained by object hallucination, i.e., generating response inconsistent with visual input. To investigate its underlying causes, we conduct an empirical analysis of internal attention distribution in various LVLMs and reveal that the bias in attention distribution imbalance varies across different architectural models. Specifically, we introduce the attention look-back rate to design a contextual entropy metric, which quantifies the balance of attention distributions across different input components (including vision, instruction and prior-generated response). This metric helps distinguish between hallucinated and non-hallucinated tokens, offering a model-agnostic method to detect over-reliance on either visual or text modalities. In light of this insight, we propose Contextual Entropy Calibration (CEC), a training-free decoding strategy that dynamically incorporates contextual entropy during beam search to steer token selection toward more balanced attention across input components. Unlike existing approaches, CEC divides inputs at each generation step into three parts and synthesizes their effects, requiring no extra training or external models, with a maximum socre improvement of 1.5% on AMBER and 79 points on MME perception Benchmark compared to previous methods. Extensive experiments on multiple hallucination benchmarks demonstrate the effectiveness of our method in mitigating object hallucinations, highlighting its potential to improve the reliability of LVLMs generation.