The rapid advancement of large language models (LLMs) has enabled natural language processing capabilities similar to those of humans, and LLMs are being widely utilized across various societal domains such as education and healthcare. While the versatility of these models has increased, they have the potential to generate subjective and normative language, leading to discriminatory treatment or outcomes among social groups, especially due to online offensive language. In this paper, we define such harm as societal bias and assess ethnic, gender, and racial biases in a model fine-tuned with Korean comments using Bidirectional Encoder Representations from Transformers (KcBERT) and KOLD data through template-based Masked Language Modeling (MLM). To quantitatively evaluate biases, we employ LPBS and CBS metrics. Compared to KcBERT, the fine-tuned model shows a reduction in ethnic bias but demonstrates significant changes in gender and racial biases. Based on these results, we propose two methods to mitigate societal bias. Firstly, a data balancing approach during the pre-training phase adjusts the uniformity of data by aligning the distribution of the occurrences of specific words and converting surrounding harmful words into non-harmful words. Secondly, during the in-training phase, we apply Debiasing Regularization by adjusting dropout and regularization, confirming a decrease in training loss. Our contribution lies in demonstrating that societal bias exists in Korean language models due to language-dependent characteristics.

In image-based robot manipulation tasks with large observation and action spaces, reinforcement learning struggles with low sample efficiency, slow training speed, and uncertain convergence. As an alternative, large pre-trained foundation models have shown promise in robotic manipulation, particularly in zero-shot and few-shot applications. However, using these models directly is unreliable due to limited reasoning capabilities and challenges in understanding physical and spatial contexts. This paper introduces ExploRLLM, a novel approach that leverages the inductive bias of foundation models (e.g. Large Language Models) to guide exploration in reinforcement learning. We also exploit these foundation models to reformulate the action and observation spaces to enhance the training efficiency in reinforcement learning. Our experiments demonstrate that guided exploration enables much quicker convergence than training without it. Additionally, we validate that ExploRLLM outperforms vanilla foundation model baselines and that the policy trained in simulation can be applied in real-world settings without additional training.

Auxiliary function is a helpful component to improve language model's code generation ability. However, a systematic exploration of how they affect has yet to be done. In this work, we comprehensively evaluate the ability to utilize auxiliary functions encoded in recent code-pretrained language models. First, we construct a human-crafted evaluation set, called HumanExtension, which contains examples of two functions where one function assists the other. With HumanExtension, we design several experiments to examine their ability in a multifaceted way. Our evaluation processes enable a comprehensive understanding of including auxiliary functions in the prompt in terms of effectiveness and robustness. An additional implementation style analysis captures the models' various implementation patterns when they access the auxiliary function. Through this analysis, we discover the models' promising ability to utilize auxiliary functions including their self-improving behavior by implementing the two functions step-by-step. However, our analysis also reveals the model's underutilized behavior to call the auxiliary function, suggesting the future direction to enhance their implementation by eliciting the auxiliary function call ability encoded in the models. We release our code and dataset to facilitate this research direction.

A promising approach for improving the performance of vision-language models like CLIP for image classification is to extend the class descriptions (i.e., prompts) with related attributes, e.g., using brown sparrow instead of sparrow. However, current zero-shot methods select a subset of attributes regardless of commonalities between the target classes, potentially providing no useful information that would have helped to distinguish between them. For instance, they may use color instead of bill shape to distinguish between sparrows and wrens, which are both brown. We propose Follow-up Differential Descriptions (FuDD), a zero-shot approach that tailors the class descriptions to each dataset and leads to additional attributes that better differentiate the target classes. FuDD first identifies the ambiguous classes for each image, and then uses a Large Language Model (LLM) to generate new class descriptions that differentiate between them. The new class descriptions resolve the initial ambiguity and help predict the correct label. In our experiments, FuDD consistently outperforms generic description ensembles and naive LLM-generated descriptions on 12 datasets. We show that differential descriptions are an effective tool to resolve class ambiguities, which otherwise significantly degrade the performance. We also show that high quality natural language class descriptions produced by FuDD result in comparable performance to few-shot adaptation methods. What is the most distinguishing characteristic of a sparrow? To distinguish it from what?

However, due to a lack of mental health awareness and fear of stigma, many patients do not actively seek diagnosis and treatment, leading to detrimental outcomes. Depression detection aims to determine whether an individual suffers from depression by analyzing their history of posts on social media, which can significantly aid in early detection and intervention. It mainly faces two key challenges: 1) it requires professional medical knowledge, and 2) it necessitates both high accuracy and explainability. To address it, we propose a novel depression detection system called DORIS, combining medical knowledge and the recent advances in large language models (LLMs). Specifically, to tackle the first challenge, we proposed an LLM-based solution to first annotate whether high-risk texts meet medical diagnostic criteria. Further, we retrieve texts with high emotional intensity and summarize critical information from the historical mood records of users, so-called mood courses. To tackle the second challenge, we combine LLM and traditional classifiers to integrate medical knowledge-guided features, for which the model can also explain its prediction results, achieving both high accuracy and explainability. Extensive experimental results on benchmarking datasets show that, compared to the current best baseline, our approach improves by 0.036 in AUPRC, which can be considered significant, demonstrating the effectiveness of our approach and its high value as an NLP application.

In this study, we tackle a growing concern around the safety and ethical use of large language models (LLMs). Despite their potential, these models can be tricked into producing harmful or unethical content through various sophisticated methods, including 'jailbreaking' techniques and targeted manipulation. Our work zeroes in on a specific issue: to what extent LLMs can be led astray by asking them to generate responses that are instruction-centric such as a pseudocode, a program or a software snippet as opposed to vanilla text. To investigate this question, we introduce TechHazardQA, a dataset containing complex queries which should be answered in both text and instruction-centric formats (e.g., pseudocodes), aimed at identifying triggers for unethical responses. We query a series of LLMs -- Llama-2-13b, Llama-2-7b, Mistral-V2 and Mistral 8X7B -- and ask them to generate both text and instruction-centric responses. For evaluation we report the harmfulness score metric as well as judgements from GPT-4 and humans. Overall, we observe that asking LLMs to produce instruction-centric responses enhances the unethical response generation by ~2-38% across the models. As an additional objective, we investigate the impact of model editing using the ROME technique, which further increases the propensity for generating undesirable content. In particular, asking edited LLMs to generate instruction-centric responses further increases the unethical response generation by ~3-16% across the different models.

Human-centered dynamic scene understanding plays a pivotal role in enhancing the capability of robotic and autonomous systems, in which Video-based Human-Object Interaction (V-HOI) detection is a crucial task in semantic scene understanding, aimed at comprehensively understanding HOI relationships within a video to benefit the behavioral decisions of mobile robots and autonomous driving systems. Although previous V-HOI detection models have made significant strides in accurate detection on specific datasets, they still lack the general reasoning ability like human beings to effectively induce HOI relationships. In this study, we propose V-HOI Multi-LLMs Collaborated Reasoning (V-HOI MLCR), a novel framework consisting of a series of plug-and-play modules that could facilitate the performance of current V-HOI detection models by leveraging the strong reasoning ability of different off-the-shelf pre-trained large language models (LLMs). We design a two-stage collaboration system of different LLMs for the V-HOI task. Specifically, in the first stage, we design a Cross-Agents Reasoning scheme to leverage the LLM conduct reasoning from different aspects. In the second stage, we perform Multi-LLMs Debate to get the final reasoning answer based on the different knowledge in different LLMs. Additionally, we devise an auxiliary training strategy that utilizes CLIP, a large vision-language model to enhance the base V-HOI models' discriminative ability to better cooperate with LLMs. We validate the superiority of our design by demonstrating its effectiveness in improving the prediction accuracy of the base V-HOI model via reasoning from multiple perspectives.

Gender bias represents a form of systematic negative treatment that targets individuals based on their gender. This discrimination can range from subtle sexist remarks and gendered stereotypes to outright hate speech. Prior research has revealed that ignoring online abuse not only affects the individuals targeted but also has broader societal implications. These consequences extend to the discouragement of women's engagement and visibility within public spheres, thereby reinforcing gender inequality. This thesis investigates the nuances of how gender bias is expressed through language and within language technologies. Significantly, this thesis expands research on gender bias to multilingual contexts, emphasising the importance of a multilingual and multicultural perspective in understanding societal biases. In this thesis, I adopt an interdisciplinary approach, bridging natural language processing with other disciplines such as political science and history, to probe gender bias in natural language and language models.

We introduce EXAMS-V, a new challenging multi-discipline multimodal multilingual exam benchmark for evaluating vision language models. It consists of 20,932 multiple-choice questions across 20 school disciplines covering natural science, social science, and other miscellaneous studies, e.g., religion, fine arts, business, etc. EXAMS-V includes a variety of multimodal features such as text, images, tables, figures, diagrams, maps, scientific symbols, and equations. The questions come in 11 languages from 7 language families. Unlike existing benchmarks, EXAMS-V is uniquely curated by gathering school exam questions from various countries, with a variety of education systems. This distinctive approach calls for intricate reasoning across diverse languages and relies on region-specific knowledge. Solving the problems in the dataset requires advanced perception and joint reasoning over the text and the visual content of the image. Our evaluation results demonstrate that this is a challenging dataset, which is difficult even for advanced vision-text models such as GPT-4V and Gemini; this underscores the inherent complexity of the dataset and its significance as a future benchmark.

We proposed an end-to-end system design towards utilizing Retrieval Augmented Generation (RAG) to improve the factual accuracy of Large Language Models (LLMs) for domain-specific and time-sensitive queries related to private knowledge-bases. Our system integrates RAG pipeline with upstream datasets processing and downstream performance evaluation. Addressing the challenge of LLM hallucinations, we finetune models with a curated dataset which originates from CMU's extensive resources and annotated with the teacher model. Our experiments demonstrate the system's effectiveness in generating more accurate answers to domain-specific and time-sensitive inquiries. The results also revealed the limitations of fine-tuning LLMs with small-scale and skewed datasets. This research highlights the potential of RAG systems in augmenting LLMs with external datasets for improved performance in knowledge-intensive tasks. Our code and models are available on Github.