Using English and Chinese multilingual and monolingual models, we analyze the internal circuitry involved in two tasks, one focusing on indirect object identification (IOI) which is virtually identical between the languages, and one which involves generating paste tense verbs that require morphological marking in English but not in Chinese. Our contributions are as follows: We show that a multilingual model uses a single circuit to handle the same syntactic process independently of the language in which it occurs ( 3.4). We show that even monolingual models trained independently on English and Chinese each adopt nearly the same circuit for this task ( 3.5), suggesting a surprising amount of consistency with how LLMs learn to handle this particular aspect of language modeling. Finally, we show that, when faced with similar tasks that require language-specific morphological processes, multilingual models still invoke a largely overlapping circuit, but employ language-specific components as needed. Specifically, in our task, we find that the model uses a circuit that consists primarily of attention heads to perform most of the task, but employs the feed-forward networks in English only to perform morphological marking that is necessary in English but not in Chinese ( 4). Together, our results provide new insights into how LLMs trade off between exploiting common structures and preserving linguistic differences when tasked with modeling multiple languages simultaneously. Our experiments can lay the groundwork for future works which seek to improve cross-lingual transfer through more principled parameter updates (Wu et al., 2024), as well as work which seeks to use LLMs in order to improve the study of linguistic and grammatical structure for its own sake (Lakretz et al., 2021; Misra & Kim, 2024).

In this paper, we propose an extension to Longformer Encoder-Decoder, a popular sparse transformer architecture. One common challenge with sparse transformers is that they can struggle with encoding of long range context, such as connections between topics discussed at a beginning and end of a document. A method to selectively increase global attention is proposed and demonstrated for abstractive summarization tasks on several benchmark data sets. By prefixing the transcript with additional keywords and encoding global attention on these keywords, improvement in zero-shot, few-shot, and fine-tuned cases is demonstrated for some benchmark data sets.

Enhancing the adaptive capabilities of large language models is a critical pursuit in both research and application. Traditional fine-tuning methods require substantial data and computational resources, especially for enhancing specific capabilities, while in-context learning is limited by the need for appropriate demonstrations and efficient token usage. Inspired by the expression of in-context learned capabilities through task vectors and the concept of modularization, we propose \alg, a framework consisting of two modules designed to effectively store and reuse task vectors to elicit the diverse capabilities of models without additional training or inference tokens. Our comprehensive experiments and analysis demonstrate that our pipeline is highly transferable across different input formats, tasks, and model architectures. ELICIT serves as a plug-and-play performance booster to enable adaptive elicitation of model capabilities. By externally storing and reusing vectors that represent in-context learned capabilities, \alg not only demonstrates the potential to operate modular capabilities but also significantly enhances the performance, versatility, adaptability, and scalability of large language models. Our code will be publicly available at https://github.com/LINs-lab/ELICIT.

Explainable molecular property prediction is essential for various scientific fields, such as drug discovery and material science. Despite delivering intrinsic explainability, linear models struggle with capturing complex, non-linear patterns. Large language models (LLMs), on the other hand, yield accurate predictions through powerful inference capabilities yet fail to provide chemically meaningful explanations for their predictions. This work proposes a novel framework, called MoleX, which leverages LLM knowledge to build a simple yet powerful linear model for accurate molecular property prediction with faithful explanations. The core of MoleX is to model complicated molecular structure-property relationships using a simple linear model, augmented by LLM knowledge and a crafted calibration strategy. Specifically, to extract the maximum amount of task-relevant knowledge from LLM embeddings, we employ information bottleneck-inspired fine-tuning and sparsity-inducing dimensionality reduction. These informative embeddings are then used to fit a linear model for explainable inference. Moreover, we introduce residual calibration to address prediction errors stemming from linear models' insufficient expressiveness of complex LLM embeddings, thus recovering the LLM's predictive power and boosting overall accuracy. Theoretically, we provide a mathematical foundation to justify MoleX's explainability. Extensive experiments demonstrate that MoleX outperforms existing methods in molecular property prediction, establishing a new milestone in predictive performance, explainability, and efficiency. In particular, MoleX enables CPU inference and accelerates large-scale dataset processing, achieving comparable performance 300x faster with 100,000 fewer parameters than LLMs. Additionally, the calibration improves model performance by up to 12.7% without compromising explainability.

Explainable Artificial Intelligence (XAI) has emerged as a critical area of research to unravel the opaque inner logic of (deep) machine learning models. Among the various XAI techniques proposed in the literature, counterfactual explanations stand out as one of the most promising approaches. However, these ``what-if'' explanations are frequently complex and technical, making them difficult for non-experts to understand and, more broadly, challenging for humans to interpret. To bridge this gap, in this work, we exploit the power of open-source Large Language Models to generate natural language explanations when prompted with valid counterfactual instances produced by state-of-the-art explainers for graph-based models. Experiments across several graph datasets and counterfactual explainers show that our approach effectively produces accurate natural language representations of counterfactual instances, as demonstrated by key performance metrics.

Multimodal attributed graphs (MAGs) are prevalent in various real-world scenarios and generally contain two kinds of knowledge: (a) Attribute knowledge is mainly supported by the attributes of different modalities contained in nodes (entities) themselves, such as texts and images. (b) Topology knowledge, on the other hand, is provided by the complex interactions posed between nodes. The cornerstone of MAG representation learning lies in the seamless integration of multimodal attributes and topology. Recent advancements in Pre-trained Language/Vision models (PLMs/PVMs) and Graph neural networks (GNNs) have facilitated effective learning on MAGs, garnering increased research interest. However, the absence of meaningful benchmark datasets and standardized evaluation procedures for MAG representation learning has impeded progress in this field. In this paper, we propose Multimodal Attribute Graph Benchmark (MAGB)}, a comprehensive and diverse collection of challenging benchmark datasets for MAGs. The MAGB datasets are notably large in scale and encompass a wide range of domains, spanning from e-commerce networks to social networks. In addition to the brand-new datasets, we conduct extensive benchmark experiments over MAGB with various learning paradigms, ranging from GNN-based and PLM-based methods, to explore the necessity and feasibility of integrating multimodal attributes and graph topology. In a nutshell, we provide an overview of the MAG datasets, standardized evaluation procedures, and present baseline experiments. The entire MAGB project is publicly accessible at https://github.com/sktsherlock/ATG.

Long-context modeling is one of the critical capabilities of language AI for digesting and reasoning over complex information pieces. In practice, long-context capabilities are typically built into a pre-trained language model (LM) through a carefully designed context extension stage, with the goal of producing generalist long-context capabilities. In our preliminary experiments, however, we discovered that the current open-weight generalist long-context models are still lacking in practical long-context processing tasks. While this means perfectly effective long-context modeling demands task-specific data, the cost can be prohibitive. In this paper, we draw inspiration from how humans process a large body of information: a lossy retrieval stage ranks a large set of documents while the reader ends up reading deeply only the top candidates. We build an automatic data synthesis pipeline that mimics this process using short-context LMs. The short-context LMs are further tuned using these self-generated data to obtain task-specific longcontext capabilities. Similar to how pre-training learns from imperfect data, we hypothesize and further demonstrate that the short-context model can bootstrap over the synthetic data, outperforming not only long-context generalist models but also the retrieval and read pipeline used to synthesize the training data in realworld tasks such as long-context retrieval augmented generation. The field of Artificial Intelligence (AI) and Natural Language Processing (NLP) have made substantial progress in building and teaching neural language models (LMs) to understand and generate language (Radford et al., 2019; Brown et al., 2020; OpenAI, 2023; Anthropic, 2023; 2024; Touvron et al., 2023a;b; MetaAI et al., 2024). Large-scale deep learning has enabled large LMs to learn from massive amounts of human-generated text (Radford et al., 2019; Brown et al., 2020).

Synthetic data has gained attention for training large language models, but poor-quality data can harm performance (see, e.g., Shumailov et al. (2023); Seddik et al. (2024)). A potential solution is data pruning, which retains only high-quality data based on a score function (human or machine feedback). Previous work Feng et al. (2024) analyzed models trained on synthetic data as sample size increases. We extend this by using random matrix theory to derive the performance of a binary classifier trained on a mix of real and pruned synthetic data in a high dimensional setting. Our findings identify conditions where synthetic data could improve performance, focusing on the quality of the generative model and verification strategy. We also show a smooth phase transition in synthetic label noise, contrasting with prior sharp behavior in infinite sample limits. Experiments with toy models and large language models validate our theoretical results.

We introduce LLMD, a large language model designed to analyze a patient's medical history based on their medical records. Along with domain knowledge, LLMD is trained on a large corpus of records collected over time and across facilities, as well as tasks and labels that make nuanced connections among them. This approach is critical to an accurate picture of patient health, and has distinctive advantages over models trained on knowledge alone, unlabeled records, structured EHR data, or records from a single health system. The recipe for LLMD continues pretraining a foundational model on both domain knowledge and the contents of millions of records. These span an average of 10 years of care and as many as 140 care sites per patient. LLMD is then instruction fine-tuned on structuring and abstraction tasks. The former jointly identify and normalize document metadata, provenance information, clinical named-entities, and ontology mappings, while the latter roll these into higher-level representations, such a continuous era of time a patient was on a medication. LLMD is deployed within a layered validation system that includes continual random audits and review by experts, e.g. based on uncertainty, disease-specific rules, or use-case. LLMD exhibits large gains over both more-powerful generalized models and domain-specific models. On medical knowledge benchmarks, LLMD-8B achieves state of the art accuracy on PubMedQA text responses, besting orders-of-magnitude larger models. On production tasks, we show that LLMD significantly outperforms all other models evaluated, and among alternatives, large general purpose LLMs like GPT-4o are more accurate than models emphasizing medical knowledge. We find strong evidence that accuracy on today's medical benchmarks is not the most significant factor when analyzing real-world patient data, an insight with implications for future medical LLMs.'

Social conventions are the foundation for social and economic life. As legions of AI agents increasingly interact with each other and with humans, their ability to form shared conventions will determine how effectively they will coordinate behaviors, integrate into society and influence it. Here, we investigate the dynamics of conventions within populations of Large Language Model (LLM) agents using simulated interactions. First, we show that globally accepted social conventions can spontaneously arise from local interactions between communicating LLMs. Second, we demonstrate how strong collective biases can emerge during this process, even when individual agents appear to be unbiased. Third, we examine how minority groups of committed LLMs can drive social change by establishing new social conventions. We show that once these minority groups reach a critical size, they can consistently overturn established behaviors. In all cases, contrasting the experimental results with predictions from a minimal multi-agent model allows us to isolate the specific role of LLM agents. Our results clarify how AI systems can autonomously develop norms without explicit programming and have implications for designing AI systems that align with human values and societal goals.