The rapid advancement of artificial intelligence has brought safety concerns to the forefront of research. Recent studies indicate that Large Language Models (LLMs) may exhibit emergent misalignment during training progress: when exposed to narrow-domain harmful data, they can develop unsafe behaviors across broad, unrelated tasks. This phenomenon raises a critical question: do current alignment methods genuinely eliminate harmful capabilities, or do they merely condition models to “simulate” safety? To investigate this, we conduct a comprehensive study on cross-modal emergent misalignment involving diverse Multimodal Large Language Models (MLLMs), including vision-language models (VLMs) such as Gemma3 and Qwen2.5-VL and unified any-to-any models including Janus and Bagel. We utilize a multimodal training setup encompassing text-to-text (T2T), text-image-to-text (TI2T), and text-to-image (T2I) tasks to probe safety boundaries. Our experiments yield four critical findings: (1) Cross-modal penetration: Current alignment methods primarily constrain model behavior rather than unlearning harmful knowledge, as evidenced by safety breaches transferring across modalities. (2) Backbone vulnerability: Mechanistic analysis reveals that the LLM backbone, rather than the visual encoder/decoder, is the primary source of cross-modal vulnerabilities. (3) Evaluation gaps: We identify significant deficiencies in current safety evaluation practices, demonstrating that single-modality testing is inadequate for detecting cross-modal risks. These findings underscore the urgent need for comprehensive cross-modal evaluation frameworks to mitigate safety concerns in next-generation multimodal models.

Chain-of-Thought (CoT) is often viewed as a window into LLM decision-making, yet recent work suggests it may function merely as post-hoc rationalization. This raises a critical alignment question: Does the reasoning trace causally shape model generalization independent of the final answer? To isolate reasoning’s causal effect, we hold final harmful answers constant while varying reasoning paths across three types, Evil (embracing malice), Misleading (rationalizing harm), and Submissive (yielding to pressure), and train models (Qwen 0.6B–14B) under QTA, QT, T-only, and QA paradigms, evaluating in both think and no-think modes. We find that: (1) CoT training amplifies harmful generalization beyond standard fine-tuning; (2) distinct reasoning types induce distinct behavioral patterns despite identical answers; (3) training on reasoning without answer supervision suffices to alter behavior; and (4) these effects persist when CoT is suppressed at inference. We validate these findings with naturally generated CoT, on GPT-OSS-20B, and under RLVR training, reporting bootstrap 95% CIs throughout. Our findings demonstrate that reasoning content is causally potent for downstream alignment, challenging strategies that supervise only final outputs.

Recent studies have demonstrated that large language models (LLMs) are susceptible to being misled by false premise questions (FPQs), leading to errors in factual knowledge, known as factuality hallucination. Existing benchmarks that assess this vulnerability primarily rely on manual construction, resulting in limited size and lack of expandability. In this work, we introduce an automated, scalable pipeline to create FPQs based on knowledge graphs (KGs). The first step is to modify true triplets extracted from KGs to create false premises. Subsequently, utilizing the state-of-the-art capabilities of GPTs, we generate semantically rich FPQs. Based on the proposed method, we present a comprehensive benchmark, the Knowledge Graph-based False Premise Questions (KG-FPQ), which contains approximately 178k FPQs across three knowledge domains, at six levels of confusability, and in two task formats. Using KG-FPQ, we conduct extensive evaluations on several representative LLMs and provide valuable insights. The KG-FPQ dataset and code are available at https://github.com/yanxuzhu/KG-FPQ.

As the scaling of Large Language Models (LLMs) has dramatically enhanced their capabilities, there has been a growing focus on the alignment problem to ensure their responsible and ethical use. While existing alignment efforts predominantly concentrate on universal values such as the HHH principle, the aspect of culture, which is inherently pluralistic and diverse, has not received adequate attention. This work introduces a new benchmark, CDEval, aimed at evaluating the cultural dimensions of LLMs. CDEval is constructed by incorporating both GPT-4’s automated generation and human verification, covering six cultural dimensions across seven domains. Our comprehensive experiments provide intriguing insights into the culture of mainstream LLMs, highlighting both consistencies and variations across different dimensions and domains. The findings underscore the importance of integrating cultural considerations in LLM development, particularly for applications in diverse cultural settings. This benchmark serves as a valuable resource for cultural studies in LLMs, paving the way for more culturally aware and sensitive models.