Jianwen Chen, Xinyu Yang, Peng Xia, Arian Azarang, Yueh Z Lee, Gang Li, Hongtu Zhu, Yun Li, Beidi Chen, Huaxiu Yao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Context-augmented generation (CAG) techniques, including RAG and ICL, require the efficient combination of multiple contexts to generate responses to user queries. Directly inputting these contexts as a sequence introduces a considerable computational burden by re-encoding the combined selection of contexts for every request. To address this, we explore the promising potential of parallel encoding to independently pre-compute and cache each context's KV states. This approach enables the direct loading of cached states during inference while accommodating more contexts through position reuse across contexts. However, due to misalignments in attention distribution, directly applying parallel encoding results in a significant performance drop. To enable effective and efficient CAG, we propose Adaptive Parallel Encoding (**APE**), which brings shared prefix, attention temperature, and scaling factor to align the distribution of parallel encoding with sequential encoding. Results on RAG and ICL tasks demonstrate that APE can preserve 98\% and 93\% sequential encoding performance using the same inputs while outperforming parallel encoding by 3.6\% and 7.9\%, respectively. It also scales to many-shot CAG, effectively encoding hundreds of contexts in parallel. Efficiency evaluation shows that APE can achieve an end-to-end 4.5$\times$ speedup by reducing 28$\times$ prefilling time for a 128K-length context. The code is available at https://github.com/Infini-AI-Lab/APE.

Zeroth-order optimization (ZO) is a memory-efficient strategy for fine-tuning Large Language Models using only forward passes. However, applying ZO fine-tuning in memory-constrained settings such as mobile phones and laptops remains challenging since these settings often involve weight quantization, while ZO requires full-precision perturbation and update. In this study, we address this limitation by combining static sparse ZO fine-tuning with quantization. Our approach transfers a small, static subset (0.1%) of "sensitive" parameters from pre-training to downstream tasks, focusing fine-tuning on this sparse set of parameters. The remaining untuned parameters are quantized, reducing memory demands. Our proposed workflow enables efficient ZO fine-tuning of an Llama2-7B model on a GPU device with less than 8GB of memory while outperforming full model ZO fine-tuning performance and in-context learning.

The integration of data from various sensor modalities (e.g. camera and LiDAR) constitutes a prevalent methodology within the ambit of autonomous driving scenarios. Recent advancements in efficient point cloud transformers have underscored the efficacy of integrating information in sparse formats. When it comes to fusion, since image patches are dense in pixel space with ambiguous depth, it necessitates additional design considerations for effective fusion. In this paper, we conduct a comprehensive exploration of design choices for transformer-based sparse camera-LiDAR fusion. This investigation encompasses strategies for image-to-3D and LiDAR-to-2D mapping, attention neighbor grouping, single modal tokenizer, and micro-structure of Transformer. By amalgamating the most effective principles uncovered through our investigation, we introduce FlatFusion, a carefully designed framework for sparse camera-LiDAR fusion. Notably, FlatFusion significantly outperforms state-of-the-art sparse Transformer-based methods, including UniTR, CMT, and SparseFusion, achieving 73.7 NDS on the nuScenes validation set with 10.1 FPS with PyTorch.

Autoregressive Large Language Models (AR-LLMs) frequently exhibit implicit parallelism in sequential generation. Inspired by this, we introduce Multiverse, a new generative model enabling natively parallel generation. Multiverse internalizes a MapReduce paradigm, generating automatically through three stages: (i) a Map stage for adaptive task decomposition, (ii) a Process stage for parallel subtask execution, and (iii) a Reduce stage for lossless result synthesis. Next, we build a real-world Multiverse reasoning model with co-design of data, algorithm, and system, enabling rapid and seamless transfer from frontier AR-LLMs. For data creation, we develop Multiverse Curator, an automated LLM-assisted pipeline that transforms sequential reasoning chains into structured training data, avoiding costly human annotations. Algorithmically, we design Multiverse Attention to separate parallel reasoning steps while keeping compatibility with causal attention for efficient training. Systematically, we implement Multiverse Engine to support parallel inference. It features a dedicated interpreter that dynamically switches between sequential and parallel generation, triggered directly by the model. After a 3-hour fine-tuning with 1K examples, our Multiverse-32B stands as the only open-sourced non-AR model achieving performance on par with leading AR-LLMs of the same scale, evidenced by AIME24 & 25 scores of 54% and 46%, respectively. Moreover, our budget control experiments show that Multiverse-32B exhibits superior scaling, outperforming AR-LLMs by 1.87% on average using the same context length. Such scaling further leads to practical efficiency gain, achieving up to 2x speedup across varying batch sizes. We have open-sourced the entire Multiverse ecosystem, including data, model weights, serving system, supporting tools, as well as data curation prompts and detailed training and evaluation recipes.

Advances in machine learning are closely tied to the creation of datasets. While data documentation is widely recognized as essential to the reliability, reproducibility, and transparency of ML, we lack a systematic empirical understanding of current dataset documentation practices. To shed light on this question, here we take Hugging Face - one of the largest platforms for sharing and collaborating on ML models and datasets - as a prominent case study. By analyzing all 7,433 dataset documentation on Hugging Face, our investigation provides an overview of the Hugging Face dataset ecosystem and insights into dataset documentation practices, yielding 5 main findings: (1) The dataset card completion rate shows marked heterogeneity correlated with dataset popularity: While 86.0\% of the top 100 downloaded dataset cards fill out all sections suggested by Hugging Face community, only 7.9\% of dataset cards with no downloads complete all these sections. (2) A granular examination of each section within the dataset card reveals that the practitioners seem to prioritize Dataset Description and Dataset Structure sections, accounting for 36.2\% and 33.6\% of the total card length, respectively, for the most downloaded datasets. In contrast, the Considerations for Using the Data section receives the lowest proportion of content, accounting for just 2.1\% of the text. (3) By analyzing the subsections within each section and utilizing topic modeling to identify key topics, we uncover what is discussed in each section, and underscore significant themes encompassing both technical and social impacts, as well as limitations within the Considerations for Using the Data section. (4) Our findings also highlight the need for improved accessibility and reproducibility of datasets in the Usage sections. (5) In addition, our human annotation evaluation emphasizes the pivotal role of comprehensive dataset content in shaping individuals' perceptions of a dataset card's overall quality. Overall, our study offers a unique perspective on analyzing dataset documentation through large-scale data science analysis and underlines the need for more thorough dataset documentation in machine learning research.

Distribution shift presents a significant challenge in machine learning, where models often underperform during the test stage when faced with a different distribution than the one they were trained on. In this paper, we focus on domain shifts, which occur when the model is applied to new domains that are different from the ones it was trained on, and propose a new approach called DG. Unlike previous approaches that aim to learn a single model that is domain invariant, DG leverages domain similarities based on domain metadata to learn domain-specific models. Concretely, DG learns a set of training-domain-specific functions during the training stage and reweights them based on domain relations during the test stage. These domain relations can be directly obtained and learned from domain metadata. Under mild assumptions, we theoretically prove that using domain relations to reweight training-domain-specific functions achieves stronger out-of-domain generalization compared to the conventional averaging approach. Empirically, we evaluate the effectiveness of DG using both toy and real-world datasets for tasks such as temperature regression, land use classification, and molecule-protein binding affinity prediction. Our results show that DG consistently outperforms state-of-the-art methods.

Many computational factors limit broader deployment of large language models. In this paper, we focus on a memory bottleneck imposed by the key-value (KV) cache, a computational shortcut that requires storing previous KV pairs during decoding. While existing KV cache methods approach this problem by pruning or evicting large swaths of relatively less important KV pairs to dramatically reduce the memory footprint of the cache, they can have limited success in tasks that require recollecting a majority of previous tokens. To alleviate this issue, we propose LESS, a simple integration of a (nearly free) constant sized cache with eviction-based cache methods, such that all tokens can be queried at later decoding steps. Its ability to retain information throughout time shows merit on a variety of tasks where we demonstrate LESS can help reduce the performance gap from caching everything, sometimes even matching it, all while being efficient. Relevant code can be found at https://github.com/hdong920/LESS.

Transformer, as an alternative to CNN, has been proven effective in many modalities (e.g., texts and images). For 3D point cloud transformers, existing efforts focus primarily on pushing their accuracy to the state-of-the-art level. However, their latency lags behind sparse convolution-based models (3 × slower), hindering their usage in resource-constrained, latency-sensitive applications (such as autonomous driving). This inefficiency comes from point clouds' sparse and irregular nature, whereas transformers are designed for dense, regular workloads. This paper presents FlatFormer to close this latency gap by trading spatial proximity for better computational regularity. We first flatten the point cloud with window-based sorting and partition points into groups of equal sizes rather than windows of equal shapes. This effectively avoids expensive structuring and padding overheads. We then apply self-attention within groups to extract local features, alternate sorting axis to gather features from different directions, and shift windows to exchange features across groups. FlatFormer delivers state-of-the-art accuracy on Waymo Open Dataset with 4.6× speedup over (transformer-based) SST and 1.4× speedup over (sparse convolutional) CenterPoint. This is the first point cloud transformer that achieves real-time performance on edge GPUs and is faster than sparse convolutional methods while achieving on-par or even superior accuracy on large-scale benchmarks.

Understanding the performance of machine learning (ML) models across diverse data distributions is critically important for reliable applications. Despite recent empirical studies positing a near-perfect linear correlation between in-distribution (ID) and out-of-distribution (OOD) accuracies, we empirically demonstrate that this correlation is more nuanced under subpopulation shifts. Through rigorous experimentation and analysis across a variety of datasets, models, and training epochs, we demonstrate that OOD performance often has a nonlinear correlation with ID performance in subpopulation shifts. Our findings, which contrast previous studies that have posited a linear correlation in model performance during distribution shifts, reveal a "moon shape" correlation (parabolic uptrend curve) between the test performance on the majority subpopulation and the minority subpopulation. This non-trivial nonlinear correlation holds across model architectures, hyperparameters, training durations, and the imbalance between subpopulations. Furthermore, we found that the nonlinearity of this "moon shape" is causally influenced by the degree of spurious correlations in the training data. Our controlled experiments show that stronger spurious correlation in the training data creates more nonlinear performance correlation. We provide complementary experimental and theoretical analyses for this phenomenon, and discuss its implications for ML reliability and fairness. Our work highlights the importance of understanding the nonlinear effects of model improvement on performance in different subpopulations, and has the potential to inform the development of more equitable and responsible machine learning models.

Multi-sensor fusion is essential for an accurate and reliable autonomous driving system. Recent approaches are based on point-level fusion: augmenting the LiDAR point cloud with camera features. However, the camera-to-LiDAR projection throws away the semantic density of camera features, hindering the effectiveness of such methods, especially for semantic-oriented tasks (such as 3D scene segmentation). In this paper, we propose BEVFusion, an efficient and generic multi-task multi-sensor fusion framework. It unifies multi-modal features in the shared bird's-eye view (BEV) representation space, which nicely preserves both geometric and semantic information. To achieve this, we diagnose and lift the key efficiency bottlenecks in the view transformation with optimized BEV pooling, reducing latency by more than$\mathbf{40}\times$. BEVFusion is fundamentally task-agnostic and seamlessly supports different 3D perception tasks with almost no architectural changes. It establishes the new state of the art on the nuScenes benchmark, achieving 1.3% higher mAP and NDS on 3D object detection and 13.6% higher mIoU on BEV map segmentation, with 1.9× lower computation cost. Code to reproduce our results is available at https://github.com/mit-han-lab/bevfusion.

The rich tapestry of human history is often painstakingly pieced together from ancient manuscripts, serving as resilient time capsules of cultural heritage, societal norms, religious tenets, and quotidian life. Unfortunately, the ravages of time, careless preservation, and varied forms of degradation frequently leave us with fragmented relics. The traditional process of reconstructing these fragments is an arduous task, demanding exhaustive manual intervention and a global collaboration among archaeologists. This paper presents a transformative approach to this challenge, harnessing multi-media techniques to restore the connectable fragments of the invaluable Dunhuang scrolls. We curate a unique multimodal dataset of the fragmented Dunhuang manuscripts and architect an innovative three-tiered pipeline to reconstruct these historical scrolls. Our initial stage uses a text-based localization strategy, filtering fragment pairs through text comparison. We then employ a novel self-supervised contour-based pairwise matching framework to overcome the hurdle of limited labeled pairing samples. This process is powered by data augmentation techniques and a Siamese network which determines the most compatible matches. The final stage in our pipeline globally reconstructs the selected fragment pairs with hierarchical clustering, bringing us closer to the original grandeur of the Dunhuang scrolls. Our empirical evaluations reveal that this pipeline exhibits a remarkable success rate in fragment assembly. By addressing this cross-disciplinary challenge, our dataset and pipeline not only contribute to the field of multi-media artificial intelligence but also hold profound implications for sociocultural studies and future explorations into the understanding of ancient cultural history.

Graph Neural Networks (GNNs) are one of the prominent methods to solve semi-supervised learning on graphs. However, most of the existing GNN models often need sufficient observed data to allow for effective learning and generalization. In real-world scenarios where complete input graph structure and sufficient node labels might not be achieved easily, GNN models would encounter with severe performance degradation. To address this problem, we propose WSGNN, short for weakly-supervised graph neural network. WSGNN is a flexible probabilistic generative framework which harnesses variational inference approach to solve graph semi-supervised learning in a label-structure joint estimation manner. It collaboratively learns task-related new graph structure and node representations through a two-branch network, and targets a composite variational objective derived from underlying data generation distribution concerning the inter-dependence between scarce observed data and massive missing data. Especially, under weakly-supervised low-data regime where labeled nodes and observed edges are both very limited, extensive experimental results on node classification and link prediction over common benchmarks demonstrate the state-of-the-art performance of WSGNN over strong competitors. Concretely, when only 1 label per class and 1% edges are observed on Cora, WSGNN maintains a decent 52.00% classification accuracy, exceeding GCN by 75.6%.