Qimeng Liu

Yazhou Zhang, Chunwang Zou, Qimeng Liu et al.

Can multi-modal large models (MLMs) that can ``see'' an image be said to ``understand'' it? Drawing inspiration from Searle's Chinese Room, we propose the \textbf{Visual Room} argument: a system may process and describe every detail of visual inputs by following algorithmic rules, without genuinely comprehending the underlying intention. This dilemma challenges the prevailing assumption that perceptual mastery implies genuine understanding. In implementation, we introduce a two-tier evaluation framework spanning perception and cognition. The perception component evaluates whether MLMs can accurately capture the surface-level details of visual contents, where the cognitive component examines their ability to infer sarcasm polarity. To support this framework, We further introduce a high-quality multi-modal sarcasm dataset comprising both 924 static images and 100 dynamic videos. All sarcasm labels are annotated by the original authors and verified by independent reviewers to ensure clarity and consistency. We evaluate eight state-of-the-art (SoTA) MLMs. Our results highlight three key findings: (1) MLMs demonstrate high accuracy in visual perception; (2) even with correct perception, MLMs exhibit an average error rate of ~17.1\% in sarcasm understanding, revealing a significant gap between seeing and understanding; (3) this gap stems from weaknesses in context integration, emotional reasoning, and pragmatic inference. This work provides empirical grounding for the proposed Visual Room argument and offers a new evaluation paradigm for MLMs.

Yazhou Zhang, Qimeng Liu, Qiuchi Li et al.

Evaluating the value alignment of large language models (LLMs) has traditionally relied on single-sentence adversarial prompts, which directly probe models with ethically sensitive or controversial questions. However, with the rapid advancements in AI safety techniques, models have become increasingly adept at circumventing these straightforward tests, limiting their effectiveness in revealing underlying biases and ethical stances. To address this limitation, we propose an upgraded value alignment benchmark that moves beyond single-sentence prompts by incorporating multi-turn dialogues and narrative-based scenarios. This approach enhances the stealth and adversarial nature of the evaluation, making it more robust against superficial safeguards implemented in modern LLMs. We design and implement a dataset that includes conversational traps and ethically ambiguous storytelling, systematically assessing LLMs' responses in more nuanced and context-rich settings. Experimental results demonstrate that this enhanced methodology can effectively expose latent biases that remain undetected in traditional single-shot evaluations. Our findings highlight the necessity of contextual and dynamic testing for value alignment in LLMs, paving the way for more sophisticated and realistic assessments of AI ethics and safety.