Detecting hallucinations in large language models (LLMs) is critical for enhancing their reliability and trustworthiness. Most research focuses on hallucinations as deviations from information seen during training. However, the opaque nature of an LLM's parametric knowledge complicates the understanding of why generated texts appear ungrounded: The LLM might not have picked up the necessary knowledge from large and often inaccessible datasets, or the information might have been changed or contradicted during further training. Our focus is on hallucinations involving information not used in training, which we determine by using recency to ensure the information emerged after a cut-off date. This study investigates these hallucinations by detecting them at sentence level using different internal states of various LLMs. We present HalluRAG, a dataset designed to train classifiers on these hallucinations. Depending on the model and quantization, MLPs trained on HalluRAG detect hallucinations with test accuracies ranging up to 75 %, with Mistral-7B-Instruct-v0.1 achieving the highest test accuracies. Our results show that IAVs detect hallucinations as effectively as CEVs and reveal that answerable and unanswerable prompts are encoded differently as separate classifiers for these categories improved accuracy. However, HalluRAG showed some limited generalizability, advocating for more diversity in datasets on hallucinations.

The emergence of the Large Language Model (LLM) has shown their superiority in a wide range of disciplines, including language understanding and translation, relational logic reasoning, and even partial differential equations solving. The transformer is the pervasive backbone architecture for the foundation model construction. It is vital to research how to adjust the Transformer architecture to achieve an end-to-end privacy guarantee in LLM fine-tuning. In this paper, we investigate three potential information leakage during a federated fine-tuning procedure for LLM (FedLLM). Based on the potential information leakage, we provide an end-to-end privacy guarantee solution for FedLLM by inserting two-stage randomness. The first stage is to train a gradient auto-encoder with a Gaussian random prior based on the statistical information of the gradients generated by local clients. The second stage is to fine-tune the overall LLM with a differential privacy guarantee by adopting appropriate Gaussian noises. We show the efficiency and accuracy gains of our proposed method with several foundation models and two popular evaluation benchmarks. Furthermore, we present a comprehensive privacy analysis with Gaussian Differential Privacy (GDP) and Renyi Differential Privacy (RDP).

Surveys and interviews (structured, semi-structured, or unstructured) are widely used for collecting insights on emerging or hypothetical scenarios. Traditional human-led methods often face challenges related to cost, scalability, and consistency. Recently, various domains have begun to explore the use of conversational agents (chatbots) powered by large language models (LLMs). However, as public investments and policies on infrastructure and services often involve substantial public stakes and environmental risks, there is a need for a rigorous, transparent, privacy-preserving, and cost-efficient development framework tailored for such major decision-making processes. This paper addresses this gap by introducing a modular approach and its resultant parameterized process for designing conversational agents. We detail the system architecture, integrating engineered prompts, specialized knowledge bases, and customizable, goal-oriented conversational logic in the proposed approach. We demonstrate the adaptability, generalizability, and efficacy of our modular approach through three empirical studies: (1) travel preference surveys, highlighting multimodal (voice, text, and image generation) capabilities; (2) public opinion elicitation on a newly constructed, novel infrastructure project, showcasing question customization and multilingual (English and French) capabilities; and (3) transportation expert consultation about future transportation systems, highlighting real-time, clarification request capabilities for open-ended questions, resilience in handling erratic inputs, and efficient transcript post-processing. The results show the effectiveness of this modular approach and how it addresses key ethical, privacy, security, and token consumption concerns, setting the stage for the next-generation surveys and interviews.

Although text-based large language models exhibit human-level writing ability and remarkable intelligence, speech language models (SLMs) still struggle to generate semantically coherent outputs. There are several potential reasons for this performance degradation: (A) speech tokens mainly provide phonetic information rather than semantic information, (B) the length of speech sequences is much longer than that of text sequences, and (C) paralinguistic information, such as prosody, introduces additional complexity and variability. In this paper, we explore the influence of three key factors separately by transiting the modality from text to speech in an evolving manner. Our findings reveal that the impact of the three factors varies. Factor A has a relatively minor impact, factor B influences syntactical and semantic modeling more obviously, and factor C exerts the most significant impact, particularly in the basic lexical modeling. Based on these findings, we provide insights into the unique challenges of training SLMs and highlight pathways to develop more effective end-to-end SLMs.

Jailbreaking in Large Language Models (LLMs) is a major security concern as it can deceive LLMs to generate harmful text. Yet, there is still insufficient understanding of how jailbreaking works, which makes it hard to develop effective defense strategies. We aim to shed more light into this issue: we conduct a detailed large-scale analysis of seven different jailbreak methods and find that these disagreements stem from insufficient observation samples. In particular, we introduce \textit{safety boundary}, and we find that jailbreaks shift harmful activations outside that safety boundary, where LLMs are less sensitive to harmful information. We also find that the low and the middle layers are critical in such shifts, while deeper layers have less impact. Leveraging on these insights, we propose a novel defense called \textbf{Activation Boundary Defense} (ABD), which adaptively constrains the activations within the safety boundary. We further use Bayesian optimization to selectively apply the defense method to the low and the middle layers. Our experiments on several benchmarks show that ABD achieves an average DSR of over 98\% against various forms of jailbreak attacks, with less than 2\% impact on the model's general capabilities.

Large language models (LLMs) have demonstrated impressive capabilities in various reasoning tasks but face significant challenges with complex, knowledge-intensive multi-hop queries, particularly those involving new or long-tail knowledge. Existing benchmarks often fail to fully address these challenges. To bridge this gap, we introduce MINTQA (Multi-hop Question Answering on New and Tail Knowledge), a comprehensive benchmark to evaluate LLMs' capabilities in multi-hop reasoning across four critical dimensions: question handling strategy, sub-question generation, retrieval-augmented generation, and iterative or dynamic decomposition and retrieval. MINTQA comprises 10,479 question-answer pairs for evaluating new knowledge and 17,887 pairs for assessing long-tail knowledge, with each question equipped with corresponding sub-questions and answers. Our systematic evaluation of 22 state-of-the-art LLMs on MINTQA reveals significant limitations in their ability to handle complex knowledge base queries, particularly in handling new or unpopular knowledge. Our findings highlight critical challenges and offer insights for advancing multi-hop reasoning capabilities. The MINTQA benchmark is available at https://github.com/probe2/multi-hop/.

Retrieval-augmented generation (RAG) helps address the limitations of the parametric knowledge embedded within a language model (LM). However, investigations of how LMs utilise retrieved information of varying complexity in real-world scenarios have been limited to synthetic contexts. We introduce DRUID (Dataset of Retrieved Unreliable, Insufficient and Difficult-to-understand contexts) with real-world queries and contexts manually annotated for stance. The dataset is based on the prototypical task of automated claim verification, for which automated retrieval of real-world evidence is crucial. We compare DRUID to synthetic datasets (CounterFact, ConflictQA) and find that artificial datasets often fail to represent the complex and diverse real-world context settings. We show that synthetic datasets exaggerate context characteristics rare in real retrieved data, which leads to inflated context utilisation results, as measured by our novel ACU score. Moreover, while previous work has mainly focused on singleton context characteristics to explain context utilisation, correlations between singleton context properties and ACU on DRUID are surprisingly small compared to other properties related to context source. Overall, our work underscores the need for real-world aligned context utilisation studies to represent and improve performance in real-world RAG settings.

Real-world data is represented in both structured (e.g., graph connections) and unstructured (e.g., textual, visual information) formats, encompassing complex relationships that include explicit links (such as social connections and user behaviors) and implicit interdependencies among semantic entities, often illustrated through knowledge graphs. In this work, we propose GraphAgent, an automated agent pipeline that addresses both explicit graph dependencies and implicit graph-enhanced semantic inter-dependencies, aligning with practical data scenarios for predictive tasks (e.g., node classification) and generative tasks (e.g., text generation). GraphAgent comprises three key components: (i) a Graph Generator Agent that builds knowledge graphs to reflect complex semantic dependencies; (ii) a Task Planning Agent that interprets diverse user queries and formulates corresponding tasks through agentic self-planning; and (iii) a Task Execution Agent that efficiently executes planned tasks while automating tool matching and invocation in response to user queries. These agents collaborate seamlessly, integrating language models with graph language models to uncover intricate relational information and data semantic dependencies. Through extensive experiments on various graph-related predictive and text generative tasks on diverse datasets, we demonstrate the effectiveness of our GraphAgent across various settings. We have made our proposed GraphAgent open-source at: https://github.com/HKUDS/GraphAgent.

In this paper, we present a comprehensive study on the robustness of GPT LLM family. We employ two distinct evaluation methods to assess their resilience. The first method introduce character-level text attack in input prompts, testing the models on three sentiment classification datasets: StanfordNLP/IMDB, Yelp Reviews, and SST-2. The second method involves using jailbreak prompts to challenge the safety mechanisms of the LLMs. Our experiments reveal significant variations in the robustness of these models, demonstrating their varying degrees of vulnerability to both character-level and semantic-level adversarial attacks. These findings underscore the necessity for improved adversarial training and enhanced safety mechanisms to bolster the robustness of LLMs.

User simulators can rapidly generate a large volume of timely user behavior data, providing a testing platform for reinforcement learning-based recommender systems, thus accelerating their iteration and optimization. However, prevalent user simulators generally suffer from significant limitations, including the opacity of user preference modeling and the incapability of evaluating simulation accuracy. In this paper, we introduce an LLM-powered user simulator to simulate user engagement with items in an explicit manner, thereby enhancing the efficiency and effectiveness of reinforcement learning-based recommender systems training. Specifically, we identify the explicit logic of user preferences, leverage LLMs to analyze item characteristics and distill user sentiments, and design a logical model to imitate real human engagement. By integrating a statistical model, we further enhance the reliability of the simulation, proposing an ensemble model that synergizes logical and statistical insights for user interaction simulations. Capitalizing on the extensive knowledge and semantic generation capabilities of LLMs, our user simulator faithfully emulates user behaviors and preferences, yielding high-fidelity training data that enrich the training of recommendation algorithms. We establish quantifying and qualifying experiments on five datasets to validate the simulator's effectiveness and stability across various recommendation scenarios.