Despite the success of Transformers on language understanding, code generation, and logical reasoning, they still fail to generalize over length on basic arithmetic tasks such as addition and multiplication. A major reason behind this failure is the vast difference in structure between numbers and text; For example, the numbers are typically parsed from right to left, and there is a correspondence between digits at the same position across different numbers. In contrast, for text, such symmetries are quite unnatural. In this work, we propose to encode these semantics explicitly into the model via modified number formatting and custom positional encodings. Empirically, our method allows a Transformer trained on numbers with at most 5-digits for addition and multiplication to generalize up to 50-digit numbers, without using additional data for longer sequences. We further demonstrate that traditional absolute positional encodings (APE) fail to generalize to longer sequences, even when trained with augmented data that captures task symmetries. To elucidate the importance of explicitly encoding structure, we prove that explicit incorporation of structure via positional encodings is necessary for out-of-distribution generalization. Finally, we pinpoint other challenges inherent to length generalization beyond capturing symmetries, in particular complexity of the underlying task, and propose changes in the training distribution to address them.

Understanding how semantic meaning is encoded in the representation spaces of large language models is a fundamental problem in interpretability. In this paper, we study the two foundational questions in this area. First, how are categorical concepts, such as {'mammal', 'bird', 'reptile', 'fish'}, represented? Second, how are hierarchical relations between concepts encoded? For example, how is the fact that 'dog' is a kind of 'mammal' encoded? We show how to extend the linear representation hypothesis to answer these questions. We find a remarkably simple structure: simple categorical concepts are represented as simplices, hierarchically related concepts are orthogonal in a sense we make precise, and (in consequence) complex concepts are represented as polytopes constructed from direct sums of simplices, reflecting the hierarchical structure. We validate these theoretical results on the Gemma large language model, estimating representations for 957 hierarchically related concepts using data from WordNet.

(Renyi Qu's Master's Thesis) Recent advancements in interpretable models for vision-language tasks have achieved competitive performance; however, their interpretability often suffers due to the reliance on unstructured text outputs from large language models (LLMs). This introduces randomness and compromises both transparency and reliability, which are essential for addressing safety issues in AI systems. We introduce \texttt{Hi-CoDe} (Hierarchical Concept Decomposition), a novel framework designed to enhance model interpretability through structured concept analysis. Our approach consists of two main components: (1) We use GPT-4 to decompose an input image into a structured hierarchy of visual concepts, thereby forming a visual concept tree. (2) We then employ an ensemble of simple linear classifiers that operate on concept-specific features derived from CLIP to perform classification. Our approach not only aligns with the performance of state-of-the-art models but also advances transparency by providing clear insights into the decision-making process and highlighting the importance of various concepts. This allows for a detailed analysis of potential failure modes and improves model compactness, therefore setting a new benchmark in interpretability without compromising the accuracy.

Harmful suicide content on the Internet is a significant risk factor inducing suicidal thoughts and behaviors among vulnerable populations. Despite global efforts, existing resources are insufficient, specifically in high-risk regions like the Republic of Korea. Current research mainly focuses on understanding negative effects of such content or suicide risk in individuals, rather than on automatically detecting the harmfulness of content. To fill this gap, we introduce a harmful suicide content detection task for classifying online suicide content into five harmfulness levels. We develop a multi-modal benchmark and a task description document in collaboration with medical professionals, and leverage large language models (LLMs) to explore efficient methods for moderating such content. Our contributions include proposing a novel detection task, a multi-modal Korean benchmark with expert annotations, and suggesting strategies using LLMs to detect illegal and harmful content. Owing to the potential harm involved, we publicize our implementations and benchmark, incorporating an ethical verification process.

Facing increasingly complex BIM authoring software and the accompanying expensive learning costs, designers often seek to interact with the software in a more intelligent and lightweight manner. They aim to automate modeling workflows, avoiding obstacles and difficulties caused by software usage, thereby focusing on the design process itself. To address this issue, we proposed an LLM-based autonomous agent framework that can function as a copilot in the BIM authoring tool, answering software usage questions, understanding the user's design intentions from natural language, and autonomously executing modeling tasks by invoking the appropriate tools. In a case study based on the BIM authoring software Vectorworks, we implemented a software prototype to integrate the proposed framework seamlessly into the BIM authoring scenario. We evaluated the planning and reasoning capabilities of different LLMs within this framework when faced with complex instructions. Our work demonstrates the significant potential of LLM-based agents in design automation and intelligent interaction.

In the recent literature, large language models (LLMs) have demonstrated great promise in healthcare, for example, closed-source models such as GPT-4 [1] and MedPalm-2 [38] have shown remarkable performance, and successfully passed the United States Medical Licensing Examination (USMLE). Concurrently, open-source models like Llama 2 have also facilitated the development of specialized language models for medicine, such as MEDITRON, PMC-LLaMA, MedAlpaca, and ChatDoctors [48, 11, 25, 6], gradually bridging the performance gap with their closed-source peers. Despite these advancements, the primary focus on English-language applications by these sophisticated medical language models has constrained their potential reach, limiting the benefits to a wider, linguistically diverse audience. In the realm of open-source multilingual Large Language Models (LLMs), exemplified by BLOOM [37] and the more recent InternLM 2 [42], a notable challenge persists despite their training on diverse multilingual corpora, that is, they exhibit unsatisfactory performance on medical queries in non-English languages, a discrepancy primarily attributed to the under-representation of medical content in these general datasets. This paper endeavors to bridge this gap by developing an open-source, multilingual language model for healthcare. As shown by Figure 1, our contribution is threefold: firstly, we gather a multilingual medical corpus designed for auto-regressive training, this aims to lay a robust foundation that accurately reflects the linguistic diversity and complexity of the medical domain; secondly, to monitor the progress, we introduce a new comprehensive multilingual medical question-answering (QA) benchmark, enabling the evaluation on multi-choice QA and rationale ability of different language models under both zero-shot and fine-tuning settings; lastly, we have tested a wide spectrum of existing language models, together with those that have undergone auto-regressive pre-training on our corpus. Through this comprehensive evaluation, we aim to provide valuable insights into the models' capabilities and fostering a deeper understanding of the intricacies involved in multilingual medical query processing. For auto-regressive training, we have developed a large-scale Multilingual Medical Corpus (MMedC), amassing over 25.5 billion medical-related tokens across six primary languages: English, Chinese, Japanese, French, Russian, and Spanish. This diverse dataset was compiled from four distinct sources: (i) we devised an automatic pipeline to filter medical-related content from the broad multilingual corpus, ensuring a focused and relevant dataset, (ii) we curated an extensive collection of medical textbooks in various languages, and converted them into texts with carefully designed pre-processing, e.g., Optical Character Recognition (OCR), heuristic data filtering, etc.

Current language models decode text token by token according to probabilistic distribution, and determining the appropriate candidates for the next token is crucial to ensure generation quality. This study introduces adaptive decoding, a mechanism that dynamically empowers language models to ascertain a sensible candidate set during generation. Specifically, we introduce an entropy-based metric called confidence and conceptualize determining the optimal candidate set as a confidence-increasing process. The rationality of including a token in the candidate set is assessed by leveraging the increment of confidence. Experimental results reveal that our method balances diversity and coherence well. The human evaluation shows that our method can generate human-preferred text. Additionally, our method can potentially improve the reasoning ability of language models.

While Transformers underpin modern large language models (LMs), there is a growing list of alternative architectures with new capabilities, promises, and tradeoffs. This makes choosing the right LM architecture challenging. Recently-proposed $\textit{hybrid architectures}$ seek a best-of-all-worlds approach that reaps the benefits of all architectures. Hybrid design is difficult for two reasons: it requires manual expert-driven search, and new hybrids must be trained from scratch. We propose $\textbf{Manticore}$, a framework that addresses these challenges. Manticore $\textit{automates the design of hybrid architectures}$ while reusing pretrained models to create $\textit{pretrained}$ hybrids. Our approach augments ideas from differentiable Neural Architecture Search (NAS) by incorporating simple projectors that translate features between pretrained blocks from different architectures. We then fine-tune hybrids that combine pretrained models from different architecture families -- such as the GPT series and Mamba -- end-to-end. With Manticore, we enable LM selection without training multiple models, the construction of pretrained hybrids from existing pretrained models, and the ability to $\textit{program}$ pretrained hybrids to have certain capabilities. Manticore hybrids outperform existing manually-designed hybrids, achieve strong performance on Long Range Arena (LRA) tasks, and can improve on pretrained transformers and state space models.

The rapid advancement of Large Language Models (LLMs) in the realm of mathematical reasoning necessitates comprehensive evaluations to gauge progress and inspire future directions. Existing assessments predominantly focus on problem-solving from the examinee perspective, overlooking a dual perspective of examiner regarding error identification and correction. From the examiner perspective, we define four evaluation tasks for error identification and correction along with a new dataset with annotated error types and steps. We also design diverse prompts to thoroughly evaluate eleven representative LLMs. Our principal findings indicate that GPT-4 outperforms all models, while open-source model LLaMA-2-7B demonstrates comparable abilities to closed-source models GPT-3.5 and Gemini Pro. Notably, calculation error proves the most challenging error type. Moreover, prompting LLMs with the error types can improve the average correction accuracy by 47.9\%. These results reveal potential directions for developing the mathematical reasoning abilities of LLMs. Our code and dataset is available on https://github.com/LittleCirc1e/EIC.

LLM agents have become increasingly sophisticated, especially in the realm of cybersecurity. Researchers have shown that LLM agents can exploit real-world vulnerabilities when given a description of the vulnerability and toy capture-the-flag problems. However, these agents still perform poorly on real-world vulnerabilities that are unknown to the agent ahead of time (zero-day vulnerabilities). In this work, we show that teams of LLM agents can exploit real-world, zero-day vulnerabilities. Prior agents struggle with exploring many different vulnerabilities and long-range planning when used alone. To resolve this, we introduce HPTSA, a system of agents with a planning agent that can launch subagents. The planning agent explores the system and determines which subagents to call, resolving long-term planning issues when trying different vulnerabilities. We construct a benchmark of 15 real-world vulnerabilities and show that our team of agents improve over prior work by up to 4.5$\times$.