The pretraining data of today's strongest language models is opaque; in particular, little is known about the proportions of various domains or languages represented. In this work, we tackle a task which we call data mixture inference, which aims to uncover the distributional make-up of training data. We introduce a novel attack based on a previously overlooked source of information: byte-pair encoding (BPE) tokenizers, used by the vast majority of modern language models. Our key insight is that the ordered list of merge rules learned by a BPE tokenizer naturally reveals information about the token frequencies in its training data. Given a tokenizer's merge list along with example data for each category of interest, we formulate a linear program that solves for the proportion of each category in the tokenizer's training set. In controlled experiments, we show that our attack recovers mixture ratios with high precision for tokenizers trained on known mixtures of natural languages, programming languages, and data sources. We then apply our approach to off-the-shelf tokenizers released with recent LMs. We confirm much publicly disclosed information about these models, and also make several new inferences: GPT-4o and Mistral NeMo's tokenizers are much more multilingual than their predecessors, training on 39% and 47% non-English language data, respectively; Llama 3 extends GPT-3.5's tokenizer primarily for multilingual (48%) use; GPT-3.5's and Claude's tokenizers are trained on predominantly code (~60%). We hope our work sheds light on current design practices for pretraining data, and inspires continued research into data mixture inference for LMs.

To protect patient safety, modern pharmaceutical complexity demands strict prescription verification. We offer a new approach - Rx Strategist - that makes use of knowledge graphs and different search strategies to enhance the power of Large Language Models (LLMs) inside an agentic framework. This multifaceted technique allows for a multi-stage LLM pipeline and reliable information retrieval from a custom-built active ingredient database. Different facets of prescription verification, such as indication, dose, and possible drug interactions, are covered in each stage of the pipeline. We alleviate the drawbacks of monolithic LLM techniques by spreading reasoning over these stages, improving correctness and reliability while reducing memory demands. Our findings demonstrate that Rx Strategist surpasses many current LLMs, achieving performance comparable to that of a highly experienced clinical pharmacist. In the complicated world of modern medications, this combination of LLMs with organized knowledge and sophisticated search methods presents a viable avenue for reducing prescription errors and enhancing patient outcomes.

The advancement of Large Language Models (LLMs) for domain applications in fields such as materials science and engineering depends on the development of fine-tuning strategies that adapt models for specialized, technical capabilities. In this work, we explore the effects of Continued Pretraining (CPT), Supervised Fine-Tuning (SFT), and various preference-based optimization approaches, including Direct Preference Optimization (DPO) and Odds Ratio Preference Optimization (ORPO), on fine-tuned LLM performance. Our analysis shows how these strategies influence model outcomes and reveals that the merging of multiple fine-tuned models can lead to the emergence of capabilities that surpass the individual contributions of the parent models. We find that model merging leads to new functionalities that neither parent model could achieve alone, leading to improved performance in domain-specific assessments. Experiments with different model architectures are presented, including Llama 3.1 8B and Mistral 7B models, where similar behaviors are observed. Exploring whether the results hold also for much smaller models, we use a tiny LLM with 1.7 billion parameters and show that very small LLMs do not necessarily feature emergent capabilities under model merging, suggesting that model scaling may be a key component. In open-ended yet consistent chat conversations between a human and AI models, our assessment reveals detailed insights into how different model variants perform and show that the smallest model achieves a high intelligence score across key criteria including reasoning depth, creativity, clarity, and quantitative precision. Other experiments include the development of image generation prompts based on disparate biological material design concepts, to create new microstructures, architectural concepts, and urban design based on biological materials-inspired construction principles.

Clinical notes contain unstructured representations of patient histories, including the relationships between medical problems and prescription drugs. To investigate the relationship between cancer drugs and their associated symptom burden, we extract structured, semantic representations of medical problem and drug information from the clinical narratives of oncology notes. We present Clinical Concept Annotations for Cancer Events and Relations (CACER), a novel corpus with fine-grained annotations for over 48,000 medical problems and drug events and 10,000 drug-problem and problem-problem relations. Leveraging CACER, we develop and evaluate transformer-based information extraction (IE) models such as BERT, Flan-T5, Llama3, and GPT-4 using fine-tuning and in-context learning (ICL). In event extraction, the fine-tuned BERT and Llama3 models achieved the highest performance at 88.2-88.0 F1, which is comparable to the inter-annotator agreement (IAA) of 88.4 F1. In relation extraction, the fine-tuned BERT, Flan-T5, and Llama3 achieved the highest performance at 61.8-65.3 F1. GPT-4 with ICL achieved the worst performance across both tasks. The fine-tuned models significantly outperformed GPT-4 in ICL, highlighting the importance of annotated training data and model optimization. Furthermore, the BERT models performed similarly to Llama3. For our task, LLMs offer no performance advantage over the smaller BERT models. The results emphasize the need for annotated training data to optimize models. Multiple fine-tuned transformer models achieved performance comparable to IAA for several extraction tasks.

Predicting the performance of LLMs on individual task instances is essential to ensure their reliability in high-stakes applications. To do so, a possibility is to evaluate the considered LLM on a set of task instances and train an assessor to predict its performance based on features of the instances. However, this approach requires evaluating each new LLM on a sufficiently large set of task instances to train an assessor specific to it. In this work, we leverage the evaluation results of previously tested LLMs to reduce the number of evaluations required to predict the performance of a new LLM. In practice, we propose to test the new LLM on a small set of reference instances and train a generic assessor which predicts the performance of the LLM on an instance based on the performance of the former on the reference set and features of the instance of interest. We conduct empirical studies on HELM-Lite and KindsOfReasoning, a collection of existing reasoning datasets that we introduce, where we evaluate all instruction-fine-tuned OpenAI models until the January 2024 version of GPT4. When predicting performance on instances with the same distribution as those used to train the generic assessor, we find this achieves performance comparable to the LLM-specific assessors trained on the full set of instances. Additionally, we find that randomly selecting the reference instances performs as well as some advanced selection methods we tested. For out of distribution, however, no clear winner emerges and the overall performance is worse, suggesting that the inherent predictability of LLMs is low.

Multimodal large language models (MLLMs) equip pre-trained large-language models (LLMs) with visual capabilities. While textual prompting in LLMs has been widely studied, visual prompting has emerged for more fine-grained and free-form visual instructions. This paper presents the first comprehensive survey on visual prompting methods in MLLMs, focusing on visual prompting, prompt generation, compositional reasoning, and prompt learning. We categorize existing visual prompts and discuss generative methods for automatic prompt annotations on the images. We also examine visual prompting methods that enable better alignment between visual encoders and backbone LLMs, concerning MLLM's visual grounding, object referring, and compositional reasoning abilities. In addition, we provide a summary of model training and in-context learning methods to improve MLLM's perception and understanding of visual prompts. This paper examines visual prompting methods developed in MLLMs and provides a vision of the future of these methods.

Isidore of Seville is credited with the adage that it is language that gives birth to a people, and not the other way around , underlining the profound role played by language in the formation of cultural and social identity. Today, of the more than 7100 languages listed, a significant number are endangered. Since the 1970s, linguists, information seekers and enthusiasts have helped develop digital resources and automatic tools to support a wide range of languages, including endangered ones. The advent of Large Language Model (LLM) technologies holds both promise and peril. They offer unprecedented possibilities for the translation and generation of content and resources, key elements in the preservation and revitalisation of languages. They also present threat of homogenisation, cultural oversimplification and the further marginalisation of already vulnerable languages. The talk this paper is based on has proposed an initiatory journey, exploring the potential paths and partnerships between technology and tradition, with a particular focus on the Occitan language. Occitan is a language from Southern France, parts of Spain and Italy that played a major cultural and economic role, particularly in the Middle Ages. It is now endangered according to UNESCO. The talk critically has examined how human expertise and artificial intelligence can work together to offer hope for preserving the linguistic diversity that forms the foundation of our global and especially our European heritage while addressing some of the ethical and practical challenges that accompany the use of these powerful technologies. This paper is based on the keynote I gave at the 46th European Conference on Information Retrieval (ECIR 2024). As an alternative to reading this paper, a video talk is available online. 1 Date: 26 March 2024.

On September 16, the company is live streaming an event called Microsoft 365 Copilot: Wave 2. Microsoft's CEO Satya Nadella and corporate vice president of AI at work, Jared Spataro, will host the event on LinkedIn (It is "your AI assistant for work," so it's a fitting platform). The stream starts at 8 AM PT/11 AM ET and is available here. Spataro first announced Microsoft 365 Copilot in early 2023 to create responses, draft presentations, and break down data -- to name a few of its uses. In the year and a half since, CoPilot has folded in Microsoft's chatbot Bing and expanded to serve entire teams, generate images, and reference multiple documents when it writes. It currently costs 360 annually per user.

We investigate whether LLMs display a well-known human cognitive bias, the attraction effect, in hiring decisions. The attraction effect occurs when the presence of an inferior candidate makes a superior candidate more appealing, increasing the likelihood of the superior candidate being chosen over a non-dominated competitor. Our study finds consistent and significant evidence of the attraction effect in GPT-3.5 and GPT-4 when they assume the role of a recruiter. Irrelevant attributes of the decoy, such as its gender, further amplify the observed bias. GPT-4 exhibits greater bias variation than GPT-3.5. Our findings remain robust even when warnings against the decoy effect are included and the recruiter role definition is varied.

Pre-training and self-training are two approaches to semi-supervised learning. The comparison between pre-training and self-training has been explored. However, the previous works led to confusing findings: self-training outperforms pre-training experienced on some tasks in computer vision, and contrarily, pre-training outperforms self-training experienced on some tasks in natural language processing, under certain conditions of incomparable settings. We propose, comparatively and exhaustively, an ensemble method to empirical study all feasible training paradigms combining pre-training, self-training, and fine-tuning within consistent foundational settings comparable to data augmentation. We conduct experiments on six datasets, four data augmentation, and imbalanced data for sentiment analysis and natural language inference tasks. Our findings confirm that the pre-training and fine-tuning paradigm yields the best overall performances. Moreover, self-training offers no additional benefits when combined with semi-supervised pre-training.