Seungjong Sun

Choongwon Kang, Seungjong Sun, Hyunmin Jun et al.

As multimodal large language models (MLLMs) have advanced to process video inputs, concerns have emerged about their potential for malicious misuse. Prior jailbreak studies have shown that safety alignment in MLLMs can be bypassed through visual inputs, yet it remains unclear which properties of video inputs induce this vulnerability. To address this gap, we introduce Multi-Clip Video (MCV) SafetyBench, a dataset of 2,920 videos designed to evaluate how the diversity of video inputs affects the vulnerability of MLLMs. Each video consists of multiple short clips depicting diverse contexts related to a harmful query. Experiments on eight representative video MLLMs show that attack success consistently increases with the number of clips. Our results further indicate that the video modality is (1) more vulnerable than the image modality, (2) more vulnerable to dynamic videos than to static videos, and (3) more vulnerable when videos contain more diverse contexts. Building on these findings, we propose a defense strategy that leverages the relative robustness of the image modality.

Seungjong Sun, Seo Yeon Baek, Jang Hyun Kim

Driven by the demand for personalized AI systems, there is growing interest in aligning the behavior of large language models (LLMs) with human traits such as personality. Previous attempts to induce personality in LLMs have shown promising results, but they struggle to capture the continuous and multidimensional nature of human traits. In this work, we propose a novel method for personality modulation in LLMs via model merging. Specifically, we construct personality vectors by subtracting the weights of a pre-trained model from those of the fine-tuned model on a given personality trait. By merging personality vectors, we enable LLMs to exhibit desired personality traits without additional training. Extensive experiments show that personality vectors enable continuous control over trait intensity and support the composition of multiple traits. Furthermore, personality vectors transfer across diverse downstream models, suggesting that they encode generalizable representations of personality. Our code is available at here.

Seungjong Sun, Eungu Lee, Dongyan Nan et al.

Large language models exhibit societal biases associated with demographic information, including race, gender, and others. Endowing such language models with personalities based on demographic data can enable generating opinions that align with those of humans. Building on this idea, we propose "random silicon sampling," a method to emulate the opinions of the human population sub-group. Our study analyzed 1) a language model that generates the survey responses that correspond with a human group based solely on its demographic distribution and 2) the applicability of our methodology across various demographic subgroups and thematic questions. Through random silicon sampling and using only group-level demographic information, we discovered that language models can generate response distributions that are remarkably similar to the actual U.S. public opinion polls. Moreover, we found that the replicability of language models varies depending on the demographic group and topic of the question, and this can be attributed to inherent societal biases in the models. Our findings demonstrate the feasibility of mirroring a group's opinion using only demographic distribution and elucidate the effect of social biases in language models on such simulations.