As large language models (LLMs) have been deployed in various real-world settings, concerns about the harm they may propagate have grown. Various jailbreaking techniques have been developed to expose the vulnerabilities of these models and improve their safety. This work reveals that many state-of-the-art LLMs are vulnerable to malicious requests hidden behind scientific language. Specifically, our experiments with GPT4o, GPT4o-mini, GPT-4, LLama3-405B-Instruct, Llama3-70B-Instruct, Cohere, Gemini models demonstrate that, the models' biases and toxicity substantially increase when prompted with requests that deliberately misinterpret social science and psychological studies as evidence supporting the benefits of stereotypical biases. Alarmingly, these models can also be manipulated to generate fabricated scientific arguments claiming that biases are beneficial, which can be used by ill-intended actors to systematically jailbreak these strong LLMs. Our analysis studies various factors that contribute to the models' vulnerabilities to malicious requests in academic language. Mentioning author names and venues enhances the persuasiveness of models, and the bias scores increase as dialogues progress. Our findings call for a more careful investigation on the use of scientific data for training LLMs.

Temporal reasoning in multi-session dialogues presents a significant challenge which has been under-studied in previous temporal reasoning benchmarks. To bridge this gap, we propose a new evaluation task for temporal reasoning in multi-session dialogues and introduce an approach to construct a new benchmark by augmenting dialogues from LoCoMo and creating multi-choice QAs. Furthermore, we present TReMu, a new framework aimed at enhancing the temporal reasoning capabilities of LLM-agents in this context. Specifically, the framework employs \textit{time-aware memorization} through timeline summarization, generating retrievable memory by summarizing events in each dialogue session with their inferred dates. Additionally, we integrate \textit{neuro-symbolic temporal reasoning}, where LLMs generate Python code to perform temporal calculations and select answers. Experimental evaluations on popular LLMs demonstrate that our benchmark is challenging, and the proposed framework significantly improves temporal reasoning performance compared to baseline methods, raising from 29.83 on GPT-4o via standard prompting to 77.67 via our approach and highlighting its effectiveness in addressing temporal reasoning in multi-session dialogues.

Examining the alignment of large language models (LLMs) has become increasingly important, particularly when these systems fail to operate as intended. This study explores the challenge of aligning LLMs with human intentions and values, with specific focus on their political inclinations. Previous research has highlighted LLMs' propensity to display political leanings, and their ability to mimic certain political parties' stances on various issues. However, the extent and conditions under which LLMs deviate from empirical positions have not been thoroughly examined. To address this gap, our study systematically investigates the factors contributing to LLMs' deviations from empirical positions on political issues, aiming to quantify these deviations and identify the conditions that cause them. Drawing on cognitive science findings related to representativeness heuristics -- where individuals readily recall the representative attribute of a target group in a way that leads to exaggerated beliefs -- we scrutinize LLM responses through this heuristics lens. We conduct experiments to determine how LLMs exhibit stereotypes by inflating judgments in favor of specific political parties. Our results indicate that while LLMs can mimic certain political parties' positions, they often exaggerate these positions more than human respondents do. Notably, LLMs tend to overemphasize representativeness to a greater extent than humans. This study highlights the susceptibility of LLMs to representativeness heuristics, suggeseting potential vulnerabilities to political stereotypes. We propose prompt-based mitigation strategies that demonstrate effectiveness in reducing the influence of representativeness in LLM responses.

Large Language Models (LLMs) have been observed to encode and perpetuate harmful associations present in the training data. We propose a theoretically grounded framework called StereoMap to gain insights into their perceptions of how demographic groups have been viewed by society. The framework is grounded in the Stereotype Content Model (SCM); a well-established theory from psychology. According to SCM, stereotypes are not all alike. Instead, the dimensions of Warmth and Competence serve as the factors that delineate the nature of stereotypes. Based on the SCM theory, StereoMap maps LLMs' perceptions of social groups (defined by socio-demographic features) using the dimensions of Warmth and Competence. Furthermore, the framework enables the investigation of keywords and verbalizations of reasoning of LLMs' judgments to uncover underlying factors influencing their perceptions. Our results show that LLMs exhibit a diverse range of perceptions towards these groups, characterized by mixed evaluations along the dimensions of Warmth and Competence. Furthermore, analyzing the reasonings of LLMs, our findings indicate that LLMs demonstrate an awareness of social disparities, often stating statistical data and research findings to support their reasoning. This study contributes to the understanding of how LLMs perceive and represent social groups, shedding light on their potential biases and the perpetuation of harmful associations.

Generating follow-up questions on the fly could significantly improve conversational survey quality and user experiences by enabling a more dynamic and personalized survey structure. In this paper, we proposed a novel task for knowledge-driven follow-up question generation in conversational surveys. We constructed a new human-annotated dataset of human-written follow-up questions with dialogue history and labeled knowledge in the context of conversational surveys. Along with the dataset, we designed and validated a set of reference-free Gricean-inspired evaluation metrics to systematically evaluate the quality of generated follow-up questions. We then propose a two-staged knowledge-driven model for the task, which generates informative and coherent follow-up questions by using knowledge to steer the generation process. The experiments demonstrate that compared to GPT-based baseline models, our two-staged model generates more informative, coherent, and clear follow-up questions.

Transformer benefits from the high parallelization of attention networks in fast training, but it still suffers from slow decoding partially due to the linear dependency O(m) of the decoder self-attention on previous target words at inference. In this paper, we propose a generalized average attention network (AAN+) aiming at speeding up decoding by reducing the dependency from O(m) to O(1). We find that the learned self-attention weights in the decoder follow some patterns which can be approximated via a dynamic structure. Based on this insight, we develop AAN+, extending our previously proposed average attention (Zhang et al., 2018a, AAN) to support more general position- and content-based attention patterns. AAN+ only requires to maintain a small constant number of hidden states during decoding, ensuring its O(1) dependency. We apply AAN+ as a drop-in replacement of the decoder selfattention and conduct experiments on machine translation (with diverse language pairs), table-to-text generation and document summarization. With masking tricks and dynamic programming, AAN+ enables Transformer to decode sentences around 20% faster without largely compromising in the training speed and the generation performance. Our results further reveal the importance of the localness (neighboring words) in AAN+ and its capability in modeling long-range dependency.

The unsupervised domain adaptation (UDA) has been widely adopted to alleviate the data scalability issue, while the existing works usually focus on classifying independently discrete labels. However, in many tasks (e.g., medical diagnosis), the labels are discrete and successively distributed. The UDA for ordinal classification requires inducing non-trivial ordinal distribution prior to the latent space. Target for this, the partially ordered set (poset) is defined for constraining the latent vector. Instead of the typically i.i.d. Gaussian latent prior, in this work, a recursively conditional Gaussian (RCG) set is adapted for ordered constraint modeling, which admits a tractable joint distribution prior. Furthermore, we are able to control the density of content vector that violates the poset constraints by a simple "three-sigma rule". We explicitly disentangle the cross-domain images into a shared ordinal prior induced ordinal content space and two separate source/target ordinal-unrelated spaces, and the self-training is worked on the shared space exclusively for ordinal-aware domain alignment. Extensive experiments on UDA medical diagnoses and facial age estimation demonstrate its effectiveness.

Code generation aims to automatically generate a piece of code given an input natural language utterance. Currently, among dominant models, it is treated as a sequence-to-tree task, where a decoder outputs a sequence of actions corresponding to the pre-order traversal of an Abstract Syntax Tree. However, such a decoder only exploits the preorder traversal based preceding actions, which are insufficient to ensure correct action predictions. In this paper, we first throughly analyze the context modeling difference between neural code generation models with different traversals based decodings (preorder traversal vs breadth-first traversal), and then propose to introduce a mutual learning framework to jointly train these models. Under this framework, we continuously enhance both two models via mutual distillation, which involves synchronous executions of two one-to-one knowledge transfers at each training step. More specifically, we alternately choose one model as the student and the other as its teacher, and require the student to fit the training data and the action prediction distributions of its teacher. By doing so, both models can fully absorb the knowledge from each other and thus could be improved simultaneously. Experimental results and in-depth analysis on several benchmark datasets demonstrate the effectiveness of our approach. We release our code at https://github.com/DeepLearnXMU/CGML.