In order to make the foundation model more efficient and effective, our idea is combining sequence transformation and state transformation. First, we prove the availability of rotary position embedding in the state space duality algorithm, which reduces the perplexity of the hybrid quadratic causal self-attention and state space duality by more than 4%, to ensure that the combining sequence transformation unifies position encoding. Second, we propose dynamic mask attention, which maintains 100% accuracy in the more challenging multi-query associative recall task, improving by more than 150% compared to quadratic causal self-attention and state space duality, to ensure that the combining sequence transformation selectively filters relevant information. Third, we design cross domain mixture of experts, which makes the computational speed of expert retrieval with more than 1024 experts 8 to 10 times faster than the mixture of experts, to ensure that the combining state transformation quickly retrieval mixture. Finally, we summarize these matrix algorithms that can form the foundation model: Wonderful Matrices, which can be a competitor to popular model architectures.
Vision-Language Models (VLMs) have shown promising capabilities in handling various multimodal tasks, yet they struggle in long-context scenarios, particularly in tasks involving videos, high-resolution images, or lengthy image-text documents. In our work, we first conduct an empirical analysis of the long-context capabilities of VLMs using our augmented long-context multimodal datasets. Our findings reveal that directly applying the positional encoding mechanism used for textual tokens to visual tokens is suboptimal, and VLM performance degrades sharply when the position encoding exceeds the model's context window. To address this, we propose Variable Visual Position Encoding (V2PE), a novel positional encoding approach that employs variable and smaller increments for visual tokens, enabling more efficient management of long multimodal sequences. Our experiments demonstrate the effectiveness of V2PE to enhances VLMs' ability to effectively understand and reason over long multimodal contexts. We further integrate V2PE with our augmented long-context multimodal datasets to fine-tune the open-source VLM, InternVL2. The fine-tuned model achieves strong performance on both standard and long-context multimodal tasks. Notably, when the sequence length of the training dataset is increased to 256K tokens, the model is capable of processing multimodal sequences up to 1M tokens, highlighting its potential for real-world long-context applications.
The rapid advancement of Multimodal Large Language Models (MLLMs) has significantly impacted various multimodal tasks. However, these models face challenges in tasks that require spatial understanding within 3D environments. Efforts to enhance MLLMs, such as incorporating point cloud features, have been made, yet a considerable gap remains between the models' learned representations and the inherent complexity of 3D scenes. This discrepancy largely stems from the training of MLLMs on predominantly 2D data, which restricts their effectiveness in comprehending 3D spaces. To address this issue, in this paper, we propose a novel generalist model, i.e., Video-3D LLM, for 3D scene understanding. By treating 3D scenes as dynamic videos and incorporating 3D position encoding into these representations, our Video-3D LLM aligns video representations with real-world spatial contexts more accurately. In addition, we have implemented a maximum coverage sampling technique to optimize the trade-off between computational cost and performance. Extensive experiments demonstrate that our model achieves state-of-the-art performance on several 3D scene understanding benchmarks, including ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, and SQA3D.
Recent text-to-image diffusion models leverage cross-attention layers, which have been effectively utilized to enhance a range of visual generative tasks. However, our understanding of cross-attention layers remains somewhat limited. In this study, we present a method for constructing Head Relevance Vectors (HRVs) that align with useful visual concepts. An HRV for a given visual concept is a vector with a length equal to the total number of cross-attention heads, where each element represents the importance of the corresponding head for the given visual concept. We develop and employ an ordered weakening analysis to demonstrate the effectiveness of HRVs as interpretable features. To demonstrate the utility of HRVs, we propose concept strengthening and concept adjusting methods and apply them to enhance three visual generative tasks. We show that misinterpretations of polysemous words in image generation can be corrected in most cases, five challenging attributes in image editing can be successfully modified, and catastrophic neglect in multi-concept generation can be mitigated. Overall, our work provides an advancement in understanding cross-attention layers and introduces new approaches for fine-controlling these layers at the head level.
Recent advances in multimodal models have demonstrated impressive capabilities in object recognition and scene understanding. However, these models often struggle with precise spatial localization - a critical capability for real-world applications. Inspired by how humans use grid-based references like chess boards and maps, we propose introducing explicit visual position encoding through a simple grid overlay approach. By adding a 9x9 black grid pattern onto input images, our method provides visual spatial guidance analogous to how positional encoding works in transformers, but in an explicit, visual form. Experiments on the COCO 2017 dataset demonstrate that our grid-based approach achieves significant improvements in localization accuracy, with a 107.4% increase in IoU (from 0.27 to 0.56) and a 194.4% improvement in GIoU (from 0.18 to 0.53) compared to baseline performance. Through attention visualization analysis, we show how this visual position encoding helps models better ground spatial relationships. Our method's simplicity and effectiveness make it particularly valuable for applications requiring accurate spatial reasoning, such as robotic manipulation, medical imaging, and autonomous navigation.
Tokenization methods like Byte-Pair Encoding (BPE) enhance computational efficiency in large language models (LLMs) but often obscure internal character structures within tokens. This limitation hinders LLMs' ability to predict precise character positions, which is crucial in tasks like Chinese Spelling Correction (CSC) where identifying the positions of misspelled characters accelerates correction processes. We propose Token Internal Position Awareness (TIPA), a method that significantly improves models' ability to capture character positions within tokens by training them on reverse character prediction tasks using the tokenizer's vocabulary. Experiments demonstrate that TIPA enhances position prediction accuracy in LLMs, enabling more precise identification of target characters in original text. Furthermore, when applied to downstream tasks that do not require exact position prediction, TIPA still boosts performance in tasks needing character-level information, validating its versatility and effectiveness.
Extending context window sizes allows large language models (LLMs) to process longer sequences and handle more complex tasks. Rotary Positional Embedding (RoPE) has become the de facto standard due to its relative positional encoding properties that benefit long-context training. However, we observe that using RoPE with BFloat16 format results in numerical issues, causing it to deviate from its intended relative positional encoding, especially in long-context scenarios. This issue arises from BFloat16's limited precision and accumulates as context length increases, with the first token contributing significantly to this problem. To address this, we develop AnchorAttention, a plug-and-play attention method that alleviates numerical issues caused by BFloat16, improves long-context capabilities, and speeds up training. AnchorAttention reduces unnecessary attention computations, maintains semantic coherence, and boosts computational efficiency by treating the first token as a shared anchor with a consistent position ID, making it visible to all documents within the training context. Experiments on three types of LLMs demonstrate that AnchorAttention significantly improves long-context performance and reduces training time by over 50\% compared to standard full attention mechanisms, while preserving the original LLM's capabilities on general tasks. Our code is available at https://github.com/haonan3/AnchorContext.
Many studies have predicted SocioEconomic Position (SEP) for aggregated spatial units such as villages using satellite data, but SEP prediction at the household level and other sources of imagery have not been yet explored. We assembled a dataset of 975 households in a semi-rural district in southern Mozambique, consisting of self-reported asset, expenditure, and income SEP data, as well as multimodal imagery including satellite images and a ground-based photograph survey of 11 household elements. We fine-tuned a convolutional neural network to extract feature vectors from the images, which we then used in regression analyzes to model household SEP using different sets of image types. The best prediction performance was found when modeling asset-based SEP using random forest models with all image types, while the performance for expenditure- and income-based SEP was lower. Using SHAP, we observed clear differences between the images with the largest positive and negative effects, as well as identified the most relevant household elements in the predictions. Finally, we fitted an additional reduced model using only the identified relevant household elements, which had an only slightly lower performance compared to models using all images. Our results show how ground-based household photographs allow to zoom in from an area-level to an individual household prediction while minimizing the data collection effort by using explainable machine learning. The developed workflow can be potentially integrated into routine household surveys, where the collected household imagery could be used for other purposes, such as refined asset characterization and environmental exposure assessment.
Non-prehensile manipulation, such as pushing objects to a desired target position, is an important skill for robots to assist humans in everyday situations. However, the task is challenging due to the large variety of objects with different and sometimes unknown physical properties, such as shape, size, mass, and friction. This can lead to the object overshooting its target position, requiring fast corrective movements of the robot around the object, especially in cases where objects need to be precisely pushed. In this paper, we improve the state-of-the-art by introducing a new memory-based vision-proprioception RL model to push objects more precisely to target positions using fewer corrective movements.
Despite the substantial progress of novel view synthesis, existing methods, either based on the Neural Radiance Fields (NeRF) or more recently 3D Gaussian Splatting (3DGS), suffer significant degradation when the input becomes sparse. Numerous efforts have been introduced to alleviate this problem, but they still struggle to synthesize satisfactory results efficiently, especially in the large scene. In this paper, we propose SCGaussian, a Structure Consistent Gaussian Splatting method using matching priors to learn 3D consistent scene structure. Considering the high interdependence of Gaussian attributes, we optimize the scene structure in two folds: rendering geometry and, more importantly, the position of Gaussian primitives, which is hard to be directly constrained in the vanilla 3DGS due to the non-structure property. To achieve this, we present a hybrid Gaussian representation. Besides the ordinary non-structure Gaussian primitives, our model also consists of ray-based Gaussian primitives that are bound to matching rays and whose optimization of their positions is restricted along the ray. Thus, we can utilize the matching correspondence to directly enforce the position of these Gaussian primitives to converge to the surface points where rays intersect. Extensive experiments on forward-facing, surrounding, and complex large scenes show the effectiveness of our approach with state-of-the-art performance and high efficiency. Code is available at https://github.com/prstrive/SCGaussian.
With the rapid development of Large Language Models (LLMs), recent studies employed LLMs as recommenders to provide personalized information services for distinct users. Despite efforts to improve the accuracy of LLM-based recommendation models, relatively little attention is paid to beyond-utility dimensions. Moreover, there are unique evaluation aspects of LLM-based recommendation models, which have been largely ignored. To bridge this gap, we explore four new evaluation dimensions and propose a multidimensional evaluation framework. The new evaluation dimensions include: 1) history length sensitivity, 2) candidate position bias, 3) generation-involved performance, and 4) hallucinations. All four dimensions have the potential to impact performance, but are largely unnecessary for consideration in traditional systems. Using this multidimensional evaluation framework, along with traditional aspects, we evaluate the performance of seven LLM-based recommenders, with three prompting strategies, comparing them with six traditional models on both ranking and re-ranking tasks on four datasets. We find that LLMs excel at handling tasks with prior knowledge and shorter input histories in the ranking setting, and perform better in the re-ranking setting, beating traditional models across multiple dimensions. However, LLMs exhibit substantial candidate position bias issues, and some models hallucinate non-existent items much more often than others. We intend our evaluation framework and observations to benefit future research on the use of LLMs as recommenders. The code and data are available at https://github.com/JiangDeccc/EvaLLMasRecommender.
Numerous industrial sectors necessitate models capable of providing robust forecasts across various horizons. Despite the recent strides in crafting specific architectures for time-series forecasting and developing pre-trained universal models, a comprehensive examination of their capability in accommodating varied-horizon forecasting during inference is still lacking. This paper bridges this gap through the design and evaluation of the Elastic Time-Series Transformer (ElasTST). The ElasTST model incorporates a non-autoregressive design with placeholders and structured self-attention masks, warranting future outputs that are invariant to adjustments in inference horizons. A tunable version of rotary position embedding is also integrated into ElasTST to capture time-series-specific periods and enhance adaptability to different horizons. Additionally, ElasTST employs a multi-scale patch design, effectively integrating both fine-grained and coarse-grained information. During the training phase, ElasTST uses a horizon reweighting strategy that approximates the effect of random sampling across multiple horizons with a single fixed horizon setting. Through comprehensive experiments and comparisons with state-of-the-art time-series architectures and contemporary foundation models, we demonstrate the efficacy of ElasTST's unique design elements. Our findings position ElasTST as a robust solution for the practical necessity of varied-horizon forecasting.
This paper tackles the challenge of real-time 3D trajectory prediction for UAVs, which is critical for applications such as aerial surveillance and defense. Existing prediction models that rely primarily on position data struggle with accuracy, especially when UAV movements fall outside the position domain used in training. Our research identifies a gap in utilizing velocity estimates, first-order dynamics, to better capture the dynamics and enhance prediction accuracy and generalizability in any position domain. To bridge this gap, we propose a new trajectory prediction method using Gated Recurrent Units (GRUs) within sequence-based neural networks. Unlike traditional methods that rely on RNNs or transformers, this approach forecasts future velocities and positions based on historical velocity data instead of positions. This is designed to enhance prediction accuracy and scalability, overcoming challenges faced by conventional models in handling complex UAV dynamics. The methodology employs both synthetic and real-world 3D UAV trajectory data, capturing a wide range of flight patterns, speeds, and agility. Synthetic data is generated using the Gazebo simulator and PX4 Autopilot, while real-world data comes from the UZH-FPV and Mid-Air drone racing datasets. The GRU-based models significantly outperform state-of-the-art RNN approaches, with a mean square error (MSE) as low as 2 x 10^-8. Overall, our findings confirm the effectiveness of incorporating velocity data in improving the accuracy of UAV trajectory predictions across both synthetic and real-world scenarios, in and out of position data distributions. Finally, we open-source our 5000 trajectories dataset and a ROS 2 package to facilitate the integration with existing ROS-based UAV systems.
This paper addresses the problem of vision-based pedestrian localization, which estimates a pedestrian's location using images and camera parameters. In practice, however, calibrated camera parameters often deviate from the ground truth, leading to inaccuracies in localization. To address this issue, we propose an anchor-based method that leverages fixed-position anchors to reduce the impact of camera parameter errors. We provide a theoretical analysis that demonstrates the robustness of our approach. Experiments conducted on simulated, real-world, and public datasets show that our method significantly improves localization accuracy and remains resilient to noise in camera parameters, compared to methods without anchors.
Visual grounding has attracted wide attention thanks to its broad application in various visual language tasks. Although visual grounding has made significant research progress, existing methods ignore the promotion effect of the association between text and image features at different hierarchies on cross-modal matching. This paper proposes a Phrase Decoupling Cross-Modal Hierarchical Matching and Progressive Position Correction Visual Grounding method. It first generates a mask through decoupled sentence phrases, and a text and image hierarchical matching mechanism is constructed, highlighting the role of association between different hierarchies in cross-modal matching. In addition, a corresponding target object position progressive correction strategy is defined based on the hierarchical matching mechanism to achieve accurate positioning for the target object described in the text. This method can continuously optimize and adjust the bounding box position of the target object as the certainty of the text description of the target object improves. This design explores the association between features at different hierarchies and highlights the role of features related to the target object and its position in target positioning. The proposed method is validated on different datasets through experiments, and its superiority is verified by the performance comparison with the state-of-the-art methods.
This paper introduces Moonshine, a family of speech recognition models optimized for live transcription and voice command processing. Moonshine is based on an encoder-decoder transformer architecture and employs Rotary Position Embedding (RoPE) instead of traditional absolute position embeddings. The model is trained on speech segments of various lengths, but without using zero-padding, leading to greater efficiency for the encoder during inference time. When benchmarked against OpenAI's Whisper tiny-en, Moonshine Tiny demonstrates a 5x reduction in compute requirements for transcribing a 10-second speech segment while incurring no increase in word error rates across standard evaluation datasets. These results highlight Moonshine's potential for real-time and resource-constrained applications.
Although existing video-based 3D human mesh recovery methods have made significant progress, simultaneously estimating human pose and shape from low-resolution image features limits their performance. These image features lack sufficient spatial information about the human body and contain various noises (e.g., background, lighting, and clothing), which often results in inaccurate pose and inconsistent motion. Inspired by the rapid advance in human pose estimation, we discover that compared to image features, skeletons inherently contain accurate human pose and motion. Therefore, we propose a novel semiAnalytical Regressor using disenTangled Skeletal representations for human mesh recovery from videos, called ARTS. Specifically, a skeleton estimation and disentanglement module is proposed to estimate the 3D skeletons from a video and decouple them into disentangled skeletal representations (i.e., joint position, bone length, and human motion). Then, to fully utilize these representations, we introduce a semi-analytical regressor to estimate the parameters of the human mesh model. The regressor consists of three modules: Temporal Inverse Kinematics (TIK), Bone-guided Shape Fitting (BSF), and Motion-Centric Refinement (MCR). TIK utilizes joint position to estimate initial pose parameters and BSF leverages bone length to regress bone-aligned shape parameters. Finally, MCR combines human motion representation with image features to refine the initial human model parameters. Extensive experiments demonstrate that our ARTS surpasses existing state-of-the-art video-based methods in both per-frame accuracy and temporal consistency on popular benchmarks: 3DPW, MPI-INF-3DHP, and Human3.6M. Code is available at https://github.com/TangTao-PKU/ARTS.
Large language models (LLMs), although having revolutionized many fields, still suffer from the challenging extrapolation problem, where the inference ability of LLMs sharply declines beyond their max training lengths. In this work, we conduct a theoretical analysis to better understand why No Position Encoding (NoPE) fails outside its effective range, as well as examining the power of Position Encoding (PE) in this context. Our findings reveal that with meticulous weave position, PE can indeed be extended beyond effective range. Our theorems establish that LLMs equipped with weave PE can achieve improved extrapolation performance without additional cost. Furthermore, we introduce a novel weave PE method, Mesa-Extrapolation, which utilizes a chunk-based triangular attention matrix and applies Stair PE to manage the final chunk. This method not only retains competitive performance but also offers substantial benefits such as significantly reduced memory demand and faster inference speed. Extensive experiments validate the effectiveness of Mesa-Extrapolation, demonstrating its potential as a scalable solution to enhancing LLMs applicative reach. Our code is available at \url{https://github.com/soacker/Mesa-Extrapolation}.
Transformers often struggle with length generalization, meaning they fail to generalize to sequences longer than those encountered during training. While arithmetic tasks are commonly used to study length generalization, certain tasks are considered notoriously difficult, e.g., multi-operand addition (requiring generalization over both the number of operands and their lengths) and multiplication (requiring generalization over both operand lengths). In this work, we achieve approximately 2-3x length generalization on both tasks, which is the first such achievement in arithmetic Transformers. We design task-specific scratchpads enabling the model to focus on a fixed number of tokens per each next-token prediction step, and apply multi-level versions of \Position Coupling (Cho et al., 2024; McLeish et al., 2024) to let Transformers know the right position to attend to. On the theory side, we prove that a 1-layer Transformer using our method can solve multi-operand addition, up to operand length and operand count that are exponential in embedding dimension.
Visual Text-to-Speech (VTTS) aims to take the environmental image as the prompt to synthesize reverberant speech for the spoken content. Previous works focus on the RGB modality for global environmental modeling, overlooking the potential of multi-source spatial knowledge like depth, speaker position, and environmental semantics. To address these issues, we propose a novel multi-source spatial knowledge understanding scheme for immersive VTTS, termed MS2KU-VTTS. Specifically, we first prioritize RGB image as the dominant source and consider depth image, speaker position knowledge from object detection, and Gemini-generated semantic captions as supplementary sources. Afterwards, we propose a serial interaction mechanism to effectively integrate both dominant and supplementary sources. The resulting multi-source knowledge is dynamically integrated based on the respective contributions of each source.This enriched interaction and integration of multi-source spatial knowledge guides the speech generation model, enhancing the immersive speech experience. Experimental results demonstrate that the MS$^2$KU-VTTS surpasses existing baselines in generating immersive speech. Demos and code are available at: https://github.com/AI-S2-Lab/MS2KU-VTTS.