Qi Wang

Ryan Wang, Akshita Bhagia, Sewon Min

Large language models are typically deployed as monolithic systems, requiring the full model even when applications need only a narrow subset of capabilities, e.g., code, math, or domain-specific knowledge. Mixture-of-Experts (MoEs) seemingly offer a potential alternative by activating only a subset of experts per input, but in practice, restricting inference to a subset of experts for a given domain leads to severe performance degradation. This limits their practicality in memory-constrained settings, especially as models grow larger and sparser. We introduce EMO, an MoE designed for modularity-the independent use and composition of expert subsets-without requiring human-defined priors. Our key idea is to encourage tokens from similar domains to rely on similar experts. Since tokens within a document often share a domain, EMO restricts them to select experts from a shared pool, while allowing different documents to use different pools. This simple constraint enables coherent expert groupings to emerge during pretraining using document boundaries alone. We pretrain a 1B-active, 14B-total EMO on 1T tokens. As a full model, it matches standard MoE performance. Crucially, it enables selective expert use: retaining only 25% (12.5%) of experts incurs just a 1% (3%) absolute drop, whereas standard MoEs break under the same setting. We further find that expert subsets in EMO specialize at semantic levels (e.g., domains such as math or code), in contrast to the low-level syntactic specialization observed in standard MoEs. Altogether, our results demonstrate a path toward modular, memory-efficient deployment of large, sparse models and open new opportunities for composable architectures.

Jingwei Sun, Jun Li, Yonghong Hao et al.

In this paper, the authors propose a new approach to solving the groundwater flow equation in the Toth basin of arbitrary top and bottom topographies using deep learning. Instead of using traditional numerical solvers, they use a DeepONet to produce the boundary-to-solution mapping. This mapping takes the geometry of the physical domain along with the boundary conditions as inputs to output the steady state solution of the groundwater flow equation. To implement the DeepONet, the authors approximate the top and bottom boundaries using truncated Fourier series or piecewise linear representations. They present two different implementations of the DeepONet: one where the Toth basin is embedded in a rectangular computational domain, and another where the Toth basin with arbitrary top and bottom boundaries is mapped into a rectangular computational domain via a nonlinear transformation. They implement the DeepONet with respect to the Dirichlet and Robin boundary condition at the top and the Neumann boundary condition at the impervious bottom boundary, respectively. Using this deep-learning enabled tool, the authors investigate the impact of surface topography on the flow pattern by both the top surface and the bottom impervious boundary with arbitrary geometries. They discover that the average slope of the top surface promotes long-distance transport, while the local curvature controls localized circulations. Additionally, they find that the slope of the bottom impervious boundary can seriously impact the long-distance transport of groundwater flows. Overall, this paper presents a new and innovative approach to solving the groundwater flow equation using deep learning, which allows for the investigation of the impact of surface topography on groundwater flow patterns.

Pu Ren, Chengping Rao, Yang Liu et al.

High-fidelity simulation of complex physical systems is exorbitantly expensive and inaccessible across spatiotemporal scales. Recently, there has been an increasing interest in leveraging deep learning to augment scientific data based on the coarse-grained simulations, which is of cheap computational expense and retains satisfactory solution accuracy. However, the major existing work focuses on data-driven approaches which rely on rich training datasets and lack sufficient physical constraints. To this end, we propose a novel and efficient spatiotemporal super-resolution framework via physics-informed learning, inspired by the independence between temporal and spatial derivatives in partial differential equations (PDEs). The general principle is to leverage the temporal interpolation for flow estimation, and then introduce convolutional-recurrent neural networks for learning temporal refinement. Furthermore, we employ the stacked residual blocks with wide activation and sub-pixel layers with pixelshuffle for spatial reconstruction, where feature extraction is conducted in a low-resolution latent space. Moreover, we consider hard imposition of boundary conditions in the network to improve reconstruction accuracy. Results demonstrate the superior effectiveness and efficiency of the proposed method compared with baseline algorithms through extensive numerical experiments.

Xiaohan Cheng, Taiyuan Mei, Yun Zi et al.

Zero sample learning is an effective method for data deficiency. The existing embedded zero sample learning methods only use the known classes to construct the embedded space, so there is an overfitting of the known classes in the testing process. This project uses category semantic similarity measures to classify multiple tags. This enables it to incorporate unknown classes that have the same meaning as currently known classes into the vector space when it is built. At the same time, most of the existing zero sample learning algorithms directly use the depth features of medical images as input, and the feature extraction process does not consider semantic information. This project intends to take ELMo-MCT as the main task and obtain multiple visual features related to the original image through self-attention mechanism. In this paper, a large number of experiments are carried out on three zero-shot learning reference datasets, and the best harmonic average accuracy is obtained compared with the most advanced algorithms.

Johnny Tian-Zheng Wei, Ameya Godbole, Mohammad Aflah Khan et al.

We present Hubble, a suite of fully open-source large language models (LLMs) for the scientific study of LLM memorization. Hubble models come in standard and perturbed variants: standard models are pretrained on a large English corpus, and perturbed models are trained in the same way but with controlled insertion of text (e.g., book passages, biographies, and test sets) designed to emulate key memorization risks. Our core release includes 8 models -- standard and perturbed models with 1B or 8B parameters, pretrained on 100B or 500B tokens -- establishing that memorization risks are determined by the frequency of sensitive data relative to size of the training corpus (i.e., a password appearing once in a smaller corpus is memorized better than the same password in a larger corpus). Our release also includes 6 perturbed models with text inserted at different pretraining phases, showing that sensitive data without continued exposure can be forgotten. These findings suggest two best practices for addressing memorization risks: to dilute sensitive data by increasing the size of the training corpus, and to order sensitive data to appear earlier in training. Beyond these general empirical findings, Hubble enables a broad range of memorization research; for example, analyzing the biographies reveals how readily different types of private information are memorized. We also demonstrate that the randomized insertions in Hubble make it an ideal testbed for membership inference and machine unlearning, and invite the community to further explore, benchmark, and build upon our work.

Haoyu He, Haozheng Luo, Yan Chen et al.

Predicting human mobility is inherently challenging due to complex long-range dependencies and multi-scale periodic behaviors. To address this, we introduce RHYTHM (Reasoning with Hierarchical Temporal Tokenization for Human Mobility), a unified framework that leverages large language models (LLMs) as general-purpose spatio-temporal predictors and trajectory reasoners. Methodologically, RHYTHM employs temporal tokenization to partition each trajectory into daily segments and encode them as discrete tokens with hierarchical attention that captures both daily and weekly dependencies, thereby quadratically reducing the sequence length while preserving cyclical information. Additionally, we enrich token representations by adding pre-computed prompt embeddings for trajectory segments and prediction targets via a frozen LLM, and feeding these combined embeddings back into the LLM backbone to capture complex interdependencies. Computationally, RHYTHM keeps the pretrained LLM backbone frozen, yielding faster training and lower memory usage. We evaluate our model against state-of-the-art methods using three real-world datasets. Notably, RHYTHM achieves a 2.4% improvement in overall accuracy, a 5.0% increase on weekends, and a 24.6% reduction in training time. Code is publicly available at https://github.com/he-h/rhythm.

Haoyu He, Haozheng Luo, Qi R. Wang

Predicting human mobility across multiple cities presents significant challenges due to the complex and diverse spatial-temporal dynamics inherent in different urban environments. In this study, we propose a robust approach to predict human mobility patterns called ST-MoE-BERT. Compared to existing methods, our approach frames the prediction task as a spatial-temporal classification problem. Our methodology integrates the Mixture-of-Experts architecture with BERT model to capture complex mobility dynamics and perform the downstream human mobility prediction task. Additionally, transfer learning is integrated to solve the challenge of data scarcity in cross-city prediction. We demonstrate the effectiveness of the proposed model on GEO-BLEU and DTW, comparing it to several state-of-the-art methods. Notably, ST-MoE-BERT achieves an average improvement of 8.29%.

Xinhua Wu, Qi R. Wang

As large language models (LLMs) are increasingly applied in areas influencing societal outcomes, it is critical to understand their tendency to perpetuate and amplify biases. This study investigates whether LLMs exhibit biases in predicting human mobility -- a fundamental human behavior -- based on race and gender. Using three prominent LLMs -- GPT-4, Gemini, and Claude -- we analyzed their predictions of visitations to points of interest (POIs) for individuals, relying on prompts that included names with and without explicit demographic details. We find that LLMs frequently reflect and amplify existing societal biases. Specifically, predictions for minority groups were disproportionately skewed, with these individuals being significantly less likely to be associated with wealth-related points of interest (POIs). Gender biases were also evident, as female individuals were consistently linked to fewer career-related POIs compared to their male counterparts. These biased associations suggest that LLMs not only mirror but also exacerbate societal stereotypes, particularly in contexts involving race and gender.

Aditya Gorla, Ryan Wang, Zhengtong Liu et al.

We present CACTI, a masked autoencoding approach for imputing tabular data that leverages the structure in missingness patterns and contextual information. Our approach employs a novel median truncated copy masking training strategy that encourages the model to learn from empirical patterns of missingness while incorporating semantic relationships between features - captured by column names and text descriptions - to better represent feature dependence. These dual sources of inductive bias enable CACTI to outperform state-of-the-art methods - an average $R^2$ gain of 7.8% over the next best method (13.4%, 6.1%, and 5.3% under missing not at random, at random and completely at random, respectively) - across a diverse range of datasets and missingness conditions. Our results highlight the value of leveraging dataset-specific contextual information and missingness patterns to enhance imputation performance.

Haoyu He, Haozheng Luo, Yan Chen et al.

We introduce RHYTHM (Reasoning with Hierarchical Temporal Tokenization for Human Mobility), a framework that leverages large language models (LLMs) as spatio-temporal predictors and trajectory reasoners. RHYTHM partitions trajectories into daily segments encoded as discrete tokens with hierarchical attention, capturing both daily and weekly dependencies while substantially reducing the sequence length. Token representations are enriched with pre-computed prompt embeddings via a frozen LLM, enhancing the model's ability to capture interdependencies without extensive computational overhead. By freezing the LLM backbone, RHYTHM achieves significant computational efficiency. Evaluation on three real-world datasets demonstrates a 2.4% improvement in accuracy, 5.0% increase on weekends, and 24.6% reduction in training time compared to state-of-the-art methods.

Ryan Wang, Matthew Finlayson, Luca Soldaini et al. · allen-ai

Language model developers typically filter out high-risk content -- such as toxic or copyrighted text -- from their pre-training data to prevent models from generating similar outputs. However, removing such data altogether limits models' ability to recognize and appropriately respond to harmful or sensitive content. In this paper, we introduce Selective Loss to Understand but Not Generate (SLUNG), a pre-training paradigm through which models learn to understand high-risk data without learning to generate it. Instead of uniformly applying the next-token prediction loss, SLUNG selectively avoids incentivizing the generation of high-risk tokens while ensuring they remain within the model's context window. As the model learns to predict low-risk tokens that follow high-risk ones, it is forced to understand the high-risk content. Through our experiments, we show that SLUNG consistently improves models' understanding of high-risk data (e.g., ability to recognize toxic content) without increasing its generation (e.g., toxicity of model responses). Overall, our SLUNG paradigm enables models to benefit from high-risk text that would otherwise be filtered out.

Huiyang Hong, Xinkai Wu, Hongyu Sun et al.

Discovering molecules with desirable molecular properties, including ADMET profiles, is of great importance in drug discovery. Existing approaches typically employ deep learning models, such as Graph Neural Networks (GNNs) and Transformers, to predict these molecular properties by learning from diverse chemical information. However, these models often fail to efficiently capture and utilize the hierarchical nature of molecular structures, and often lack mechanisms for effective interaction among multi-level features. To address these limitations, we propose a Hierarchical Interaction Message Passing Mechanism, which serves as the foundation of our novel model, the Hierarchical Interaction Message Net (HimNet). Our method enables interaction-aware representation learning across atomic, motif, and molecular levels via hierarchical attention-guided message passing. This design allows HimNet to effectively balance global and local information, ensuring rich and task-relevant feature extraction for downstream property prediction tasks, such as Blood-Brain Barrier Permeability (BBBP). We systematically evaluate HimNet on eleven datasets, including eight widely-used MoleculeNet benchmarks and three challenging, high-value datasets for metabolic stability, malaria activity, and liver microsomal clearance, covering a broad range of pharmacologically relevant properties. Extensive experiments demonstrate that HimNet achieves the best or near-best performance in most molecular property prediction tasks. Furthermore, our method exhibits promising hierarchical interpretability, aligning well with chemical intuition on representative molecules. We believe that HimNet offers an accurate and efficient solution for molecular activity and ADMET property prediction, contributing to advanced decision-making in the early stages of drug discovery.