As the parameter scale of large language models (LLMs) grows, jointly training knowledge graph (KG) embeddings with model parameters to enhance LLM capabilities becomes increasingly costly. Consequently, the community has shown interest in developing prompt strategies that effectively integrate KG information into LLMs. However, the format for incorporating KGs into LLMs lacks standardization; for instance, KGs can be transformed into linearized triples or natural language (NL) text. Current prompting methods often rely on a trial-and-error approach, leaving researchers with an incomplete understanding of which KG input format best facilitates LLM comprehension of KG content. To elucidate this, we design a series of experiments to explore LLMs' understanding of different KG input formats within the context of prompt engineering. Our analysis examines both literal and attention distribution levels. Through extensive experiments, we indicate a counter-intuitive phenomenon: when addressing fact-related questions, unordered linearized triples are more effective for LLMs' understanding of KGs compared to fluent NL text. Furthermore, noisy, incomplete, or marginally relevant subgraphs can still enhance LLM performance. Finally, different LLMs have distinct preferences for different formats of organizing unordered triples.

The Retrieval Augmented Generation (RAG) framework utilizes a combination of parametric knowledge and external knowledge to demonstrate state-of-the-art performance on open-domain question answering tasks. However, the RAG framework suffers from performance degradation when the query is accompanied by irrelevant contexts. In this work, we propose the RE-RAG framework, which introduces a relevance estimator (RE) that not only provides relative relevance between contexts as previous rerankers did, but also provides confidence, which can be used to classify whether given context is useful for answering the given question. We propose a weakly supervised method for training the RE simply utilizing question-answer data without any labels for correct contexts. We show that RE trained with a small generator (sLM) can not only improve the sLM fine-tuned together with RE but also improve previously unreferenced large language models (LLMs). Furthermore, we investigate new decoding strategies that utilize the proposed confidence measured by RE such as choosing to let the user know that it is "unanswerable" to answer the question given the retrieved contexts or choosing to rely on LLM's parametric knowledge rather than unrelated contexts.

Puns play a vital role in academic research due to their distinct structure and clear definition, which aid in the comprehensive analysis of linguistic humor. However, the understanding of puns in large language models (LLMs) has not been thoroughly examined, limiting their use in creative writing and humor creation. In this paper, we leverage three popular tasks, i.e., pun recognition, explanation and generation to systematically evaluate the capabilities of LLMs in pun understanding. In addition to adopting the automated evaluation metrics from prior research, we introduce new evaluation methods and metrics that are better suited to the in-context learning paradigm of LLMs. These new metrics offer a more rigorous assessment of an LLM's ability to understand puns and align more closely with human cognition than previous metrics. Our findings reveal the "lazy pun generation" pattern and identify the primary challenges LLMs encounter in understanding puns.

Large Language Models (LLMs) need to be aligned with human expectations to ensure their safety and utility in most applications. Alignment is challenging, costly, and needs to be repeated for every LLM and alignment criterion. We propose to decouple LLMs and alignment by training aligner models that can be used to align any LLM for a given criteria on an as-needed basis, thus also reducing the potential negative impacts of alignment on performance. Our recipe for training the aligner models solely relies on synthetic data generated with a (prompted) LLM and can be easily adjusted for a variety of alignment criteria. We use the same synthetic data to train inspectors, binary miss-alignment classification models to guide a "squad" of multiple aligners. Our empirical results demonstrate consistent improvements when applying aligner squad to various LLMs, including chat-aligned models, across several instruction-following and red-teaming datasets.

The widespread use of social media necessitates reliable and efficient detection of offensive content to mitigate harmful effects. Although sophisticated models perform well on individual datasets, they often fail to generalize due to varying definitions and labeling of "offensive content." In this paper, we introduce HateCOT, an English dataset with over 52,000 samples from diverse sources, featuring explanations generated by GPT-3.5Turbo and curated by humans. We demonstrate that pretraining on HateCOT significantly enhances the performance of open-source Large Language Models on three benchmark datasets for offensive content detection in both zero-shot and few-shot settings, despite differences in domain and task. Additionally, HateCOT facilitates effective K-shot fine-tuning of LLMs with limited data and improves the quality of their explanations, as confirmed by our human evaluation.

The tendency of Large Language Models (LLMs) to generate hallucinations and exhibit overconfidence in predictions raises concerns regarding their reliability. Confidence or uncertainty estimations indicating the extent of trustworthiness of a model's response are essential to developing reliable AI systems. Current research primarily focuses on LLM confidence estimations in English, remaining a void for other widely used languages and impeding the global development of reliable AI applications. This paper introduces a comprehensive investigation of \textbf Multi\textbf{ling}ual \textbf{Conf}idence estimation (\textsc{MlingConf}) on LLMs. First, we introduce an elaborated and expert-checked multilingual QA dataset. Subsequently, we delve into the performance of several confidence estimation methods across diverse languages and examine how these confidence scores can enhance LLM performance through self-refinement. Extensive experiments conducted on the multilingual QA dataset demonstrate that confidence estimation results vary in different languages, and the verbalized numerical confidence estimation method exhibits the best performance among most languages over other methods. Finally, the obtained confidence scores can consistently improve performance as self-refinement feedback across various languages.

Knowledge distillation (KD) is a common approach to compress a teacher model to reduce its inference cost and memory footprint, by training a smaller student model. However, in the context of autoregressive language models (LMs), we empirically find that larger teacher LMs might dramatically result in a poorer student. In response to this problem, we conduct a series of analyses and reveal that different tokens have different teaching modes, neglecting which will lead to performance degradation. Motivated by this, we propose a simple yet effective adaptive teaching approach (ATKD) to improve the KD. The core of ATKD is to reduce rote learning and make teaching more diverse and flexible. Extensive experiments on 8 LM tasks show that, with the help of ATKD, various baseline KD methods can achieve consistent and significant performance gains (up to +3.04% average score) across all model types and sizes. More encouragingly, ATKD can improve the student model generalization effectively.

In this paper, we highlight the critical issues of robustness and safety associated with integrating large language models (LLMs) and vision-language models (VLMs) into robotics applications. Recent works focus on using LLMs and VLMs to improve the performance of robotics tasks, such as manipulation and navigation. Despite these improvements, analyzing the safety of such systems remains underexplored yet extremely critical. LLMs and VLMs are highly susceptible to adversarial inputs, prompting a significant inquiry into the safety of robotic systems. This concern is important because robotics operate in the physical world where erroneous actions can result in severe consequences. This paper explores this issue thoroughly, presenting a mathematical formulation of potential attacks on LLM/VLM-based robotic systems and offering experimental evidence of the safety challenges. Our empirical findings highlight a significant vulnerability: simple modifications to the input can drastically reduce system effectiveness. Specifically, our results demonstrate an average performance deterioration of 19.4% under minor input prompt modifications and a more alarming 29.1% under slight perceptual changes. These findings underscore the urgent need for robust countermeasures to ensure the safe and reliable deployment of advanced LLM/VLM-based robotic systems.

Kullback-Leiber divergence has been widely used in Knowledge Distillation (KD) to compress Large Language Models (LLMs). Contrary to prior assertions that reverse Kullback-Leibler (RKL) divergence is mode-seeking and thus preferable over the mean-seeking forward Kullback-Leibler (FKL) divergence, this study empirically and theoretically demonstrates that neither mode-seeking nor mean-seeking properties manifest in KD for LLMs. Instead, RKL and FKL are found to share the same optimization objective and both converge after a sufficient number of epochs. However, due to practical constraints, LLMs are seldom trained for such an extensive number of epochs. Meanwhile, we further find that RKL focuses on the tail part of the distributions, while FKL focuses on the head part at the beginning epochs. Consequently, we propose a simple yet effective Adaptive Kullback-Leiber (AKL) divergence method, which adaptively allocates weights to combine FKL and RKL. Metric-based and GPT-4-based evaluations demonstrate that the proposed AKL outperforms the baselines across various tasks and improves the diversity and quality of generated responses.

Indirect User Requests (IURs), such as "It's cold in here" instead of "Could you please increase the temperature?" are common in human-human task-oriented dialogue and require world knowledge and pragmatic reasoning from the listener. While large language models (LLMs) can handle these requests effectively, smaller models deployed on virtual assistants often struggle due to resource constraints. Moreover, existing task-oriented dialogue benchmarks lack sufficient examples of complex discourse phenomena such as indirectness. To address this, we propose a set of linguistic criteria along with an LLM-based pipeline for generating realistic IURs to test natural language understanding (NLU) and dialogue state tracking (DST) models before deployment in a new domain. We also release IndirectRequests, a dataset of IURs based on the Schema Guided Dialog (SGD) corpus, as a comparative testbed for evaluating the performance of smaller models in handling indirect requests.