Yuqing Yang, Qi Zhu, Zhen Han, Boran Han, Zhengyuan Shen, Shuai Wang, Vassilis N. Ioannidis, Huzefa Rangwala. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

We study the well-motivated problem of online distribution shift in which the data arrive in batches and the distribution of each batch can change arbitrarily over time. Since the shifts can be large or small, abrupt or gradual, the length of the relevant historical data to learn from may vary over time, which poses a major challenge in designing algorithms that can automatically adapt to the best "attention span” while remaining computationally efficient. We propose a meta-algorithm that takes any network architecture and any Online Learner (OL) algorithm as input and produces a new algorithm which provably enhances the performance of the given OL under non-stationarity. Our algorithm is efficient (it requires maintaining only $O(\log T)$ OL instances) and adaptive (it automatically chooses OL instances with the ideal "attention” length at every timestamp). Experiments on various real-world datasets across text and image modalities show that our method consistently improves the accuracy of user specified OL algorithms for classification tasks. Key novel algorithmic ingredients include a multi-resolution instance design inspired by wavelet theory and a cross-validation-through-time technique. Both could be of independent interest.

Generative models that satisfy hard constraints are critical in many scientific and engineering applications, where physical laws or system requirements must be strictly respected. Many existing constrained generative models, especially those developed for computer vision, rely heavily on gradient information, which is often sparse or computationally expensive in some fields, e.g., partial differential equations (PDEs). In this work, we introduce a novel framework for adapting pre-trained, unconstrained flow-matching models to satisfy constraints exactly in a zero-shot manner without requiring expensive gradient computations or fine-tuning. Our framework, *ECI sampling*, alternates between extrapolation (E), correction (C), and interpolation (I) stages during each iterative sampling step of flow matching sampling to ensure accurate integration of constraint information while preserving the validity of the generation. We demonstrate the effectiveness of our approach across various PDE systems, showing that ECI-guided generation strictly adheres to physical constraints and accurately captures complex distribution shifts induced by these constraints. Empirical results demonstrate that our framework consistently outperforms baseline approaches in various zero-shot constrained generation tasks and also achieves competitive results in the regression tasks without additional fine-tuning.

How to best develop foundational models for time series forecasting remains an important open question. Tokenization is a crucial consideration in this effort: what is an effective discrete vocabulary for a real-valued sequential input? To address this question, we develop WaveToken, a wavelet-based tokenizer that allows models to learn complex representations directly in the space of time-localized frequencies. Our method first scales and decomposes the input time series, then thresholds and quantizes the wavelet coefficients, and finally pre-trains an autoregressive model to forecast coefficients for the forecast horizon. By decomposing coarse and fine structures in the inputs, wavelets provide an eloquent and compact language for time series forecasting that simplifies learning. Empirical results on a comprehensive benchmark, including 42 datasets for both in-domain and zero-shot settings, show that WaveToken: i) performs on par or better than recently proposed foundation models for forecasting while using a much smaller vocabulary (1024 tokens), and is competitive with modern deep learning models trained specifically on each dataset; ii) exhibits superior generalization capabilities, achieving the best average rank across all datasets for three complementary metrics; and iii) easily captures complex temporal patterns of practical relevance that are challenging for other recent pre-trained models, including trends, sparse spikes, and non-stationary time series with varying frequencies evolving over time.

Retrieval-augmented generation (RAG) can enhance the generation quality of large language models (LLMs) by incorporating external token databases. However, retrievals from large databases can constitute a substantial portion of the overall generation time, particularly when retrievals are periodically performed to align the retrieved content with the latest states of generation. In this paper, we introduce PipeRAG, a novel algorithm-system co-design approach to reduce generation latency and enhance generation quality. PipeRAG integrates (1) pipeline parallelism to enable concurrent retrieval and generation processes, (2) flexible retrieval intervals to maximize the efficiency of pipeline parallelism, and (3) a performance model to automatically balance retrieval quality and latency based on the generation states and underlying hardware. Our evaluation shows that, by combining the three aforementioned methods, PipeRAG achieves up to 2.6× speedup in end-to-end generation latency while improving generation quality. These promising results showcase the effectiveness of co-designing algorithms with underlying systems, paving the way for the adoption of PipeRAG in future RAG systems.

Existing AutoML systems have advanced the automation of machine learning (ML); however, they still require substantial manual configuration and expert input, particularly when handling multimodal data. We introduce MLZero, a novel multi-agent framework powered by Large Language Models (LLMs) that enables end-to-end ML automation across diverse data modalities with minimal human intervention. A cognitive perception module is first employed, transforming raw multimodal inputs into perceptual context that effectively guides the subsequent workflow. To address key limitations of LLMs, such as hallucinated code generation and outdated API knowledge, we enhance the iterative code generation process with semantic and episodic memory. MLZero demonstrates superior performance on MLE-Bench Lite, outperforming all competitors in both success rate and solution quality, securing six gold medals. Furthermore, when evaluated on our Multimodal AutoML Agent Benchmark, which includes 25 more challenging tasks spanning diverse data modalities, MLZero outperforms the competing methods by a large margin with a success rate of 0.92 (+263.6\%) and an average rank of 2.28. Our approach maintains its robust effectiveness even with a compact 8B LLM, outperforming full-size systems from existing solutions.

Since the seminal work of TabPFN, research on tabular foundation models (TFMs) based on in-context learning (ICL) has challenged long-standing paradigms in machine learning. Without seeing any real-world data, models pretrained on purely synthetic datasets generalize remarkably well across diverse datasets, often using only a moderate number of in-context examples. This shifts the focus in tabular machine learning from model architecture design to the design of synthetic datasets, or, more precisely, to the prior distributions that generate them. Yet the guiding principles for prior design remain poorly understood. This work marks the first attempt to address the gap. We systematically investigate and identify key properties of synthetic priors that allow pretrained TFMs to generalize well. Based on these insights, we introduce Mitra, a TFM trained on a curated mixture of synthetic priors selected for their diversity, distinctiveness, and performance on real-world tabular data. Mitra consistently outperforms state-of-the-art TFMs, such as TabPFNv2 and TabICL, across both classification and regression benchmarks, with better sample efficiency.

In this work, we introduce Context-Aware MultiModal Learner (CaMML), for tuning large multimodal models (LMMs).CaMML, a lightweight module, is crafted to seamlessly integrate multimodal contextual samples into large models, thereby empowering the model to derive knowledge from analogous, domain-specific, up-to-date information and make grounded inferences.Importantly, CaMML is highly scalable and can efficiently handle lengthy multimodal context examples owing to its hierarchical design.Based on CaMML, we have developed two multimodal models, CaMML-7B and CaMML-13B, that have shown exceptional performance across an array of benchmark datasets for multimodal tasks.Remarkably, CaMML-13B achieves the state-of-the-art performance on over ten widely recognized multimodal benchmark datasets, surpassing LLaVA-1.5 (13B) with a noticeable margin, without integration of any external resources.Moreover, we have conducted extensive ablative studies to inspect the inner workings of CaMML and performed qualitative analyses to showcase its effectiveness in handling real-world challenging cases.Code and models are available at: https://github. com/amazon-science/camml.

In the realm of geospatial analysis, the diversity of remote sensors, encompassing both optical and microwave technologies, offers a wealth of distinct observational capabilities. Recognizing this, we present msGFM, a multisensor geospatial foundation model that effectively unifies data from four key sensor modalities. This integration spans an expansive dataset of two million multisensor images. ms-GFM is uniquely adept at handling both paired and unpaired sensor data. For data originating from identical geolocations, our model employs an innovative cross-sensor pretraining approach in masked image modeling, enabling the synthesis of joint representations from diverse sensors. ms-GFM, incorporating four remote sensors, upholds strong performance, forming a comprehensive model adaptable to various sensor types. msGFM has demonstrated enhanced proficiency in a range of both single-sensor and multisensor downstream tasks. These include scene classification, segmentation, cloud removal, and pan-sharpening. A key discovery of our research is that representations derived from natural images are not always compatible with the distinct characteristics of geospatial remote sensors, under-scoring the limitations of existing representations in this field. Our work can serve as a guide for developing multisensor geospatial pretraining models, paving the way for more advanced geospatial capabilities. Code can be found at https://github.com/boranhan/Geospatial_Foundation_Models

The question-answering (QA) capabilities of foundation models are highly sensitive to prompt variations, rendering their performance susceptible to superficial, non-meaning-altering changes. This vulnerability often stems from the model's preference or bias towards specific input characteristics, such as option position or superficial image features in multi-modal settings. We propose to rectify this bias directly in the model's internal representation. Our approach, SteerFair, finds the bias direction in the model's representation space and steers activation values away from it during inference. Specifically, we exploit the observation that bias often adheres to simple association rules, such as the spurious association between the first option and correctness likelihood. Next, we construct demonstrations of these rules from unlabeled samples and use them to identify the bias directions. We empirically show that SteerFair significantly reduces instruction-tuned model performance variance across prompt modifications on three benchmark tasks. Remarkably, our approach surpasses a supervised baseline with 100 labels by an average of 10.86% accuracy points and 12.95 score points and matches the performance with 500 labels.

How do we transfer the relevant knowledge from ever larger foundation models into small, task-specific downstream models that can run at much lower costs? Standard transfer learning using pre-trained weights as the initialization transfers limited information and commits us to often massive pre-trained architectures. This procedure also precludes combining multiple pre-trained models that learn complementary information. To address these shortcomings, we introduce Adaptive Feature Transfer (AFT). Instead of transferring weights, AFT operates purely on features, thereby decoupling the choice of the pre-trained model from the smaller downstream model. Rather than indiscriminately compressing all pre-trained features, AFT adaptively transfers pre-trained features that are most useful for performing the downstream task, using a simple regularization that adds minimal overhead. Across multiple vision, language, and multi-modal datasets, AFT achieves significantly better downstream performance compared to alternatives with a similar computational cost. Furthermore, AFT reliably translates improvement in pre-trained models into improvement in downstream performance, even if the downstream model is over $50\times$ smaller, and can effectively transfer complementary information learned by multiple pre-trained models.

While the Transformer architecture has achieved remarkable success across various domains, a thorough theoretical foundation explaining its optimization dynamics is yet to be fully developed. In this study, we aim to bridge this understanding gap by answering the following two core questions: (1) Which types of Transformer architectures allow Gradient Descent (GD) to achieve guaranteed convergence? and (2) Under what initial conditions and architectural specifics does the Transformer achieve rapid convergence during training? By analyzing the loss landscape of a single Transformer layer using Softmax and Gaussian attention kernels, our work provides concrete answers to these questions. Our findings demonstrate that, with appropriate weight initialization, GD can train a Transformer model (with either kernel type) to achieve a global optimal solution, especially when the input embedding dimension is large. Nonetheless, certain scenarios highlight potential pitfalls: training a Transformer using the Softmax attention kernel may sometimes lead to suboptimal local solutions. In contrast, the Gaussian attention kernel exhibits a much favorable behavior. Our empirical study further validate the theoretical findings.

Data augmentation is vital for object detection tasks that require expensive bounding box annotations. Recent successes in diffusion models have inspired the use of diffusion-based synthetic images for data augmentation. However, existing works have primarily focused on image classification, and their applicability to boost object detection’s performance remains unclear. To address this gap, we propose a data augmentation pipeline based on controllable diffusion models and CLIP. Our approach involves generating appropriate visual priors to control the generation of synthetic data and implementing post-filtering techniques using category-calibrated CLIP scores. The evaluation of our approach is conducted under few-shot settings in MSCOCO, full PASCAL VOC dataset, and selected downstream datasets. We observe the performance increase using our augmentation pipeline. Specifically, the mAP improvement is +18.0%/+15.6%/+15.9% for COCO 5/10/30-shot, +2.9% on full PASCAL VOC dataset, and +12.4% on average for selected downstream datasets.

Geospatial technologies are becoming increasingly essential in our world for a wide range of applications, including agriculture, urban planning, and disaster response. To help improve the applicability and performance of deep learning models on these geospatial tasks, various works have begun investigating foundation models for this domain. Researchers have explored two prominent approaches for introducing such models in geospatial applications, but both have drawbacks in terms of limited performance benefit or prohibitive training cost. Therefore, in this work, we propose a novel paradigm for building highly effective geospatial foundation models with minimal resource cost and carbon impact. We first construct a compact yet diverse dataset from multiple sources to promote feature diversity, which we term GeoPile. Then, we investigate the potential of continual pretraining from large-scale ImageNet-22k models and propose a multi-objective continual pretraining paradigm, which leverages the strong representations of ImageNet while simultaneously providing the freedom to learn valuable in-domain features. Our approach outperforms previous state-of-the-art geospatial pretraining methods in an extensive evaluation on seven downstream datasets covering various tasks such as change detection, classification, multi-label classification, semantic segmentation, and super-resolution. Code is available at https://github.com/mmendiet/GFM.

Addressing imbalanced or long-tailed data is a major challenge in visual recognition tasks due to disparities between training and testing distributions and issues with data noise. We propose the Wrapped Cauchy Distributed Angular Softmax (WCDAS), a novel softmax function that incorporates data-wise Gaussian-based kernels into the angular correlation between feature representations and classifier weights, effectively mitigating noise and sparse sampling concerns. The class-wise distribution of angular representation becomes a sum of these kernels. Our theoretical analysis reveals that the wrapped Cauchy distribution excels the Gaussian distribution in approximating mixed distributions. Additionally, WCDAS uses trainable concentration parameters to dynamically adjust the compactness and margin of each class. Empirical results confirm label-aware behavior in these parameters and demonstrate WCDAS’s superiority over other state-of-the-art softmax-based methods in handling long-tailed visual recognition across multiple benchmark datasets. The code is public available.

Earth system forecasting has traditionally relied on complex physical models that are computationally expensive and require significant domain expertise. In the past decade, the unprecedented increase in spatiotemporal Earth observation data has enabled data-driven forecasting models using deep learning techniques. These models have shown promise for diverse Earth system forecasting tasks but either struggle with handling uncertainty or neglect domain-specific prior knowledge, resulting in averaging possible futures to blurred forecasts or generating physically implausible predictions. To address these limitations, we propose a two-stage pipeline for probabilistic spatiotemporal forecasting: 1) We develop *PreDiff*, a conditional latent diffusion model capable of probabilistic forecasts. 2) We incorporate an explicit knowledge alignment mechanism to align forecasts with domain-specific physical constraints. This is achieved by estimating the deviation from imposed constraints at each denoising step and adjusting the transition distribution accordingly. We conduct empirical studies on two datasets: N-body MNIST, a synthetic dataset with chaotic behavior, and SEVIR, a real-world precipitation nowcasting dataset. Specifically, we impose the law of conservation of energy in N-body MNIST and anticipated precipitation intensity in SEVIR. Experiments demonstrate the effectiveness of PreDiff in handling uncertainty, incorporating domain-specific prior knowledge, and generating forecasts that exhibit high operational utility.

Microscopy imaging is vital in biology research and diagnosis. When imaging at the scale of cell or molecule level, mechanical drift on the axial axis can be difficult to correct. Although multi-scale networks have been developed for de-blurring, those cascade residual learning approaches fail to accurately capture the end-to-end non-linearity of deconvolution, a relation between in-focus images and their out-of-focus counterparts in microscopy. In our model, we adopt a structure of multi-scale U-Net without cascade residual leaning. Additionally, in contrast to the conventional coarse-to-fine model, our model strengthens the cross-scale interaction by fusing the features from the coarser sub-networks with the finer ones in a head-to-tail manner: the decoder from the coarser scale is fused with the encoder of the finer ones. Such interaction contributes to better feature learning as fusion happens across decoder and encoder at all scales. Numerous experiments demonstrate that our method yields better performance when compared with other existing models.