Normally, Microsoft's Build is a straightforward (often dry) showcase of the company's software and hardware developments, with a dash of on-stage coding to excite the developer crowd. But this year, the company is likely to make some huge AI moves, following its 2023 debut of ChatGPT-powered Bing Chat. In fact, Microsoft has a showcase for new Surfaces and AI in Windows 11 on May 20, while Build actual kicks off a day later. The Surface event might be the most impactful. Rumors suggest we'll see some of the first systems with Qualcomm's Arm-based Snapdragon X Elite chip alongside new features in the next major Windows 11 update. A refresh for its consumer PCs is likely to consist of new 13- and 15-inch Surface Laptop 6 models with thinner bezels, larger trackpads, improved port selection and that X Elite chip.

Large language models (LLMs) have achieved remarkable performance on a wide range of tasks. However, recent studies have shown that LLMs can memorize training data and simple repeated tokens can trick the model to leak the data. In this paper, we take a step further and show that certain special characters or their combinations with English letters are stronger memory triggers, leading to more severe data leakage. The intuition is that, since LLMs are trained with massive data that contains a substantial amount of special characters (e.g. structural symbols {, } of JSON files, and @, # in emails and online posts), the model may memorize the co-occurrence between these special characters and the raw texts. This motivates us to propose a simple but effective Special Characters Attack (SCA) to induce training data leakage. Our experiments verify the high effectiveness of SCA against state-of-the-art LLMs: they can leak diverse training data, such as code corpus, web pages, and personally identifiable information, and sometimes generate non-stop outputs as a byproduct. We further show that the composition of the training data corpus can be revealed by inspecting the leaked data -- one crucial piece of information for pre-training high-performance LLMs. Our work can help understand the sensitivity of LLMs to special characters and identify potential areas for improvement.

The addressee estimation (understanding to whom somebody is talking) is a fundamental task for human activity recognition in multi-party conversation scenarios. Specifically, in the field of human-robot interaction, it becomes even more crucial to enable social robots to participate in such interactive contexts. However, it is usually implemented as a binary classification task, restricting the robot's capability to estimate whether it was addressed and limiting its interactive skills. For a social robot to gain the trust of humans, it is also important to manifest a certain level of transparency and explainability. Explainable artificial intelligence thus plays a significant role in the current machine learning applications and models, to provide explanations for their decisions besides excellent performance. In our work, we a) present an addressee estimation model with improved performance in comparison with the previous SOTA; b) further modify this model to include inherently explainable attention-based segments; c) implement the explainable addressee estimation as part of a modular cognitive architecture for multi-party conversation in an iCub robot; d) propose several ways to incorporate explainability and transparency in the aforementioned architecture; and e) perform a pilot user study to analyze the effect of various explanations on how human participants perceive the robot.

This study investigates human and ChatGPT text simplification and its relationship to dependency distance. A set of 220 sentences, with increasing grammatical difficulty as measured in a prior user study, were simplified by a human expert and using ChatGPT. We found that the three sentence sets all differed in mean dependency distances: the highest in the original sentence set, followed by ChatGPT simplified sentences, and the human simplified sentences showed the lowest mean dependency distance. Introduction Enhancing the understandability of biomedical information is vital in fostering health-literate patients. However, empirical evidence shows that readability formulas are not appropriate tools [1], [2].

Recent advancements in large language models (LLMs) have showcased significant improvements in mathematics. However, traditional math benchmarks like GSM8k offer a unidimensional perspective, falling short in providing a holistic assessment of the LLMs' math capabilities. To address this gap, we introduce MathBench, a new benchmark that rigorously assesses the mathematical capabilities of large language models. MathBench spans a wide range of mathematical disciplines, offering a detailed evaluation of both theoretical understanding and practical problem-solving skills. The benchmark progresses through five distinct stages, from basic arithmetic to college mathematics, and is structured to evaluate models at various depths of knowledge. Each stage includes theoretical questions and application problems, allowing us to measure a model's mathematical proficiency and its ability to apply concepts in practical scenarios. MathBench aims to enhance the evaluation of LLMs' mathematical abilities, providing a nuanced view of their knowledge understanding levels and problem solving skills in a bilingual context. The project is released at https://github.com/open-compass/MathBench .

This paper details the efforts of the WisPerMed team in the BioLaySumm2024 Shared Task on automatic lay summarization in the biomedical domain, aimed at making scientific publications accessible to non-specialists. Large language models (LLMs), specifically the BioMistral and Llama3 models, were fine-tuned and employed to create lay summaries from complex scientific texts. The summarization performance was enhanced through various approaches, including instruction tuning, few-shot learning, and prompt variations tailored to incorporate specific context information. The experiments demonstrated that fine-tuning generally led to the best performance across most evaluated metrics. Few-shot learning notably improved the models' ability to generate relevant and factually accurate texts, particularly when using a well-crafted prompt. Additionally, a Dynamic Expert Selection (DES) mechanism to optimize the selection of text outputs based on readability and factuality metrics was developed. Out of 54 participants, the WisPerMed team reached the 4th place, measured by readability, factuality, and relevance. Determined by the overall score, our approach improved upon the baseline by approx. 5.5 percentage points and was only approx 1.5 percentage points behind the first place.

Current language models are built on the so-called distributional semantic approach to linguistic meaning that has the distributional hypothesis at its core. The distributional hypothesis involves a holistic conception of word meaning: the meaning of a word depends upon its relations to other words in the model. A standard objection to meaning holism is the charge of instability: any change in the meaning properties of a linguistic system (a human speaker, for example) would lead to many changes or possibly a complete change in the entire system. When the systems in question are trying to communicate with each other, it has been argued that instability of this kind makes communication impossible (Fodor and Lepore 1992, 1996, 1999). In this article, we examine whether the instability objection poses a problem for distributional models of meaning. First, we distinguish between distinct forms of instability that these models could exhibit, and we argue that only one such form is relevant for understanding the relation between instability and communication: what we call differential instability. Differential instability is variation in the relative distances between points in a space, rather than variation in the absolute position of those points. We distinguish differential and absolute instability by constructing two of our own models, a toy model constructed from the text of two novels, and a more sophisticated model constructed using the Word2vec algorithm from a combination of Wikipedia and SEP articles. We demonstrate the two forms of instability by showing how these models change as the corpora they are constructed from increase in size.

Math world problems correction(MWPC) is a novel task dedicated to rectifying reasoning errors in the process of solving mathematical problems. In this paper, leveraging the advancements in large language models (LLMs), we address two key objectives:(1) Distinguishing between mathematical reasoning and error correction; (2) Exploring strategies to enhance the error correction capabilities of LLMs in mathematics to solve MWPC task. We noticed that, in real-time education,assisting students in recognizing their mistakes is more crucial than simply providing correct answers. However, current research tends to prioritize obtaining accurate solutions to math problems rather than correcting potentially incorrect ones. Therefore, we modify the research paradigm, demonstrating that improving mathematical reasoning abilities does not equate to mastery in error correction. Meanwhile, we propose a novel method called diagnostic-oriented promping(DOP) aimed at facilitating LLMs to excel in error correction. In experiments, DOP has shown outstanding performance, highlighting its significant impact. We argue that in mathematical education, the demand for outstanding correctors surpasses that for proficient reasoners. Codes and data are available on https://github.com/ChenhaoEcnuCS/Reason-Correct.

Research on jailbreaking has been valuable for testing and understanding the safety and security issues of large language models (LLMs). In this paper, we introduce Iterative Refinement Induced Self-Jailbreak (IRIS), a novel approach that leverages the reflective capabilities of LLMs for jailbreaking with only black-box access. Unlike previous methods, IRIS simplifies the jailbreaking process by using a single model as both the attacker and target. This method first iteratively refines adversarial prompts through self-explanation, which is crucial for ensuring that even well-aligned LLMs obey adversarial instructions. IRIS then rates and enhances the output given the refined prompt to increase its harmfulness. We find IRIS achieves jailbreak success rates of 98% on GPT-4 and 92% on GPT-4 Turbo in under 7 queries. It significantly outperforms prior approaches in automatic, black-box and interpretable jailbreaking, while requiring substantially fewer queries, thereby establishing a new standard for interpretable jailbreaking methods.

Preference datasets are essential for incorporating human preferences into pre-trained language models, playing a key role in the success of Reinforcement Learning from Human Feedback. However, these datasets often demonstrate conflicting alignment objectives, leading to increased vulnerability to jailbreak attacks and challenges in adapting downstream tasks to prioritize specific alignment objectives without negatively impacting others. In this work, we introduce a novel statistical metric, Alignment Dimension Conflict, to quantify the degree of conflict within preference datasets. We then present \texttt{Hummer} and its fine-grained variant, \texttt{Hummer-F}, as innovative pairwise preference datasets with reduced-conflict alignment objectives. \texttt{Hummer} is built based on UltraFeedback and is enhanced by AI feedback from GPT-4, marking as the first preference dataset aimed at reducing the competition between alignment objectives. Furthermore, we develop reward models, HummerRM and HummerRM-F, which employ a hybrid sampling approach to balance diverse alignment objectives effectively. This sampling method positions HummerRM as an ideal model for domain-specific further fine-tuning and reducing vulnerabilities to attacks.