This paper investigates the suitability of frontier Large Language Models (LLMs) for Q&A interactions in science centres, with the aim of boosting visitor engagement while maintaining factual accuracy. Using a dataset of questions collected from the National Space Centre in Leicester (UK), we evaluated responses generated by three leading models: OpenAI's GPT-4, Claude 3.5 Sonnet, and Google Gemini 1.5. Each model was prompted for both standard and creative responses tailored to an 8-year-old audience, and these responses were assessed by space science experts based on accuracy, engagement, clarity, novelty, and deviation from expected answers. The results revealed a trade-off between creativity and accuracy, with Claude outperforming GPT and Gemini in both maintaining clarity and engaging young audiences, even when asked to generate more creative responses. Nonetheless, experts observed that higher novelty was generally associated with reduced factual reliability across all models. This study highlights the potential of LLMs in educational settings, emphasizing the need for careful prompt engineering to balance engagement with scientific rigor.

Although recent quantized Large Language Models (LLMs), such as BitNet, have paved the way for significant reduction in memory usage during deployment with binary or ternary weights, training these models still demands substantial memory footprints. This is partly because high-precision (i.e., unquantized) weight matrices required for straight-through estimation must be maintained throughout the whole training process. To address this, we explore the potential of directly updating the quantized low-precision weight matrices without relying on the straight-through estimator during backpropagation, thereby saving memory usage during training. Specifically, we employ a stochastic rounding technique to minimize information loss caused by the use of low-bit weights throughout training. Experimental results on our LLaMA-structured models indicate that (1) training with only low-precision weights is feasible even when they are constrained to ternary values, (2) extending the bit width to 8 bits results in only a 5% loss degradation compared to BitNet b1.58 while offering the potential for reduced memory usage during training, and (3) our models can also perform inference using ternary weights, showcasing their flexibility in deployment.

We consider the inference for the ranking of large language models (LLMs). Alignment arises as a big challenge to mitigate hallucinations in the use of LLMs. Ranking LLMs has been shown as a well-performing tool to improve alignment based on the best-of-$N$ policy. In this paper, we propose a new inferential framework for testing hypotheses and constructing confidence intervals of the ranking of language models. We consider the widely adopted Bradley-Terry-Luce (BTL) model, where each item is assigned a positive preference score that determines its pairwise comparisons' outcomes. We further extend it into the contextual setting, where the score of each model varies with the prompt. We show the convergence rate of our estimator. By extending the current Gaussian multiplier bootstrap theory to accommodate the supremum of not identically distributed empirical processes, we construct the confidence interval for ranking and propose a valid testing procedure. We also introduce the confidence diagram as a global ranking property. We conduct numerical experiments to assess the performance of our method.

Alignment is a social phenomenon wherein individuals share a common goal or perspective. Mirroring, or mimicking the behaviors and opinions of another individual, is one mechanism by which individuals can become aligned. Large scale investigations of the effect of mirroring on alignment have been limited due to the scalability of traditional experimental designs in sociology. In this paper, we introduce a simple computational framework that enables studying the effect of mirroring behavior on alignment in multi-agent systems. We simulate systems of interacting large language models in this framework and characterize overall system behavior and alignment with quantitative measures of agent dynamics. We find that system behavior is strongly influenced by the range of communication of each agent and that these effects are exacerbated by increased rates of mirroring. We discuss the observed simulated system behavior in the context of known human social dynamics.

Code large language models (codeLLMs) have made significant strides in code generation. Most previous code-related benchmarks, which consist of various programming exercises along with the corresponding test cases, are used as a common measure to evaluate the performance and capabilities of code LLMs. However, the current code LLMs focus on synthesizing the correct code snippet, ignoring the alignment with human preferences, where the query should be sampled from the practical application scenarios and the model-generated responses should satisfy the human preference. To bridge the gap between the model-generated response and human preference, we present a rigorous human-curated benchmark CodeArena to emulate the complexity and diversity of real-world coding tasks, where 397 high-quality samples spanning 40 categories and 44 programming languages, carefully curated from user queries. Further, we propose a diverse synthetic instruction corpus SynCode-Instruct (nearly 20B tokens) by scaling instructions from the website to verify the effectiveness of the large-scale synthetic instruction fine-tuning, where Qwen2.5-SynCoder totally trained on synthetic instruction data can achieve top-tier performance of open-source code LLMs. The results find performance differences between execution-based benchmarks and CodeArena. Our systematic experiments of CodeArena on 40+ LLMs reveal a notable performance gap between open SOTA code LLMs (e.g. Qwen2.5-Coder) and proprietary LLMs (e.g., OpenAI o1), underscoring the importance of the human preference alignment.\footnote{\url{https://codearenaeval.github.io/ }}

The detection of sensitive content in large datasets is crucial for ensuring that shared and analysed data is free from harmful material. However, current moderation tools, such as external APIs, suffer from limitations in customisation, accuracy across diverse sensitive categories, and privacy concerns. Additionally, existing datasets and open-source models focus predominantly on toxic language, leaving gaps in detecting other sensitive categories such as substance abuse or self-harm. In this paper, we put forward a unified dataset tailored for social media content moderation across six sensitive categories: conflictual language, profanity, sexually explicit material, drug-related content, self-harm, and spam. By collecting and annotating data with consistent retrieval strategies and guidelines, we address the shortcomings of previous focalised research. Our analysis demonstrates that fine-tuning large language models (LLMs) on this novel dataset yields significant improvements in detection performance compared to open off-the-shelf models such as LLaMA, and even proprietary OpenAI models, which underperform by 10-15% overall. This limitation is even more pronounced on popular moderation APIs, which cannot be easily tailored to specific sensitive content categories, among others.

The integration of Retrieval-Augmented Generation (RAG) with Multimodal Large Language Models (MLLMs) has revolutionized information retrieval and expanded the practical applications of AI. However, current systems struggle in accurately interpreting user intent, employing diverse retrieval strategies, and effectively filtering unintended or inappropriate responses, limiting their effectiveness. This paper introduces Contextual Understanding and Enhanced Search with MLLM (CUE-M), a novel multimodal search framework that addresses these challenges through a multi-stage pipeline comprising image context enrichment, intent refinement, contextual query generation, external API integration, and relevance-based filtering. CUE-M incorporates a robust filtering pipeline combining image-based, text-based, and multimodal classifiers, dynamically adapting to instance- and category-specific concern defined by organizational policies. Evaluations on a multimodal Q&A dataset and a public safety benchmark demonstrate that CUE-M outperforms baselines in accuracy, knowledge integration, and safety, advancing the capabilities of multimodal retrieval systems.

Large Audio-Language Models (LALMs) have unclocked audio dialogue capabilities, where audio dialogues are a direct exchange of spoken language between LALMs and humans. Recent advances, such as GPT-4o, have enabled LALMs in back-and-forth audio dialogues with humans. This progression not only underscores the potential of LALMs but also broadens their applicability across a wide range of practical scenarios supported by audio dialogues. However, given these advancements, a comprehensive benchmark to evaluate the performance of LALMs in the open-ended audio dialogue understanding remains absent currently. To address this gap, we propose an Audio Dialogue Understanding Benchmark (ADU-Bench), which consists of 4 benchmark datasets. They assess the open-ended audio dialogue ability for LALMs in 3 general scenarios, 12 skills, 9 multilingual languages, and 4 categories of ambiguity handling. Notably, we firstly propose the evaluation of ambiguity handling in audio dialogues that expresses different intentions beyond the same literal meaning of sentences, e.g., "Really!?" with different intonations. In summary, ADU-Bench includes over 20,000 open-ended audio dialogues for the assessment of LALMs. Through extensive experiments conducted on 13 LALMs, our analysis reveals that there is still considerable room for improvement in the audio dialogue understanding abilities of existing LALMs. In particular, they struggle with mathematical symbols and formulas, understanding human behavior such as roleplay, comprehending multiple languages, and handling audio dialogue ambiguities from different phonetic elements, such as intonations, pause positions, and homophones.

Computational argumentation, which involves generating answers or summaries for controversial topics like abortion bans and vaccination, has become increasingly important in today's polarized environment. Sophisticated LLM capabilities offer the potential to provide nuanced, evidence-based answers to such questions through Retrieval-Augmented Argumentation (RAArg), leveraging real-world evidence for high-quality, grounded arguments. However, evaluating RAArg remains challenging, as human evaluation is costly and difficult for complex, lengthy answers on complicated topics. At the same time, re-using existing argumentation datasets is no longer sufficient, as they lack long, complex arguments and realistic evidence from potentially misleading sources, limiting holistic evaluation of retrieval effectiveness and argument quality. To address these gaps, we investigate automated evaluation methods using multiple fine-grained LLM judges, providing better and more interpretable assessments than traditional single-score metrics and even previously reported human crowdsourcing. To validate the proposed techniques, we introduce ConQRet, a new benchmark featuring long and complex human-authored arguments on debated topics, grounded in real-world websites, allowing an exhaustive evaluation across retrieval effectiveness, argument quality, and groundedness. We validate our LLM Judges on a prior dataset and the new ConQRet benchmark. Our proposed LLM Judges and the ConQRet benchmark can enable rapid progress in computational argumentation and can be naturally extended to other complex retrieval-augmented generation tasks.

Despite extensive safety measures, LLMs are vulnerable to adversarial inputs, or jailbreaks, which can elicit unsafe behaviors. In this work, we introduce bijection learning, a powerful attack algorithm which automatically fuzzes LLMs for safety vulnerabilities using randomly-generated encodings whose complexity can be tightly controlled. We leverage in-context learning to teach models bijective encodings, pass encoded queries to the model to bypass built-in safety mechanisms, and finally decode responses back into English. Our attack is extremely effective on a wide range of frontier language models. Moreover, by controlling complexity parameters such as number of key-value mappings in the encodings, we find a close relationship between the capability level of the attacked LLM and the average complexity of the most effective bijection attacks. Our work highlights that new vulnerabilities in frontier models can emerge with scale: more capable models are more severely jailbroken by bijection attacks. Figure 1: An overview of the bijection learning attack, which uses in-context learning and bijective mappings with complexity parameters to optimally jailbreak LLMs of different capability levels. Large language models (LLMs) have been the subject of concerns about safety and potential misuse. As systems like Claude and ChatGPT see widespread deployment, their strong world knowledge and reasoning capabilities may empower bad actors or amplify negative side-effects of downstream usage. Hence, model designers have prudently worked to safeguard models against harmful use. Model designers have trained LLMs to refuse harmful inputs with RLHF (Christiano et al., 2017; Bai et al., 2022) and adversarial training (Ziegler et al., 2022), and have guarded LLM systems However, prior work has shown that safeguards on language models can be circumvented through adversarial input schemes. One topic of debate is how models' vulnerability to jailbreaks changes with scale. Some works (Ren et al., 2024; Howe et al., 2024) argue that scaling model capabilities can improve performance on safety benchmarks, even if defense mechanisms do not meaningfully improve.