Designing efficient optimizers for large language models (LLMs) with low-memory requirements and fast convergence is an important and challenging problem. This paper makes a step towards the systematic design of such optimizers through the lens of structured Fisher information matrix (FIM) approximation. We show that many state-of-the-art efficient optimizers can be viewed as solutions to FIM approximation (under the Frobenius norm) with specific structural assumptions. Building on these insights, we propose two design recommendations of practical efficient optimizers for LLMs, involving the careful selection of structural assumptions to balance generality and efficiency, and enhancing memory efficiency of optimizers with general structures through a novel low-rank extension framework. We demonstrate how to use each design approach by deriving new memory-efficient optimizers: Row and Column Scaled SGD (RACS) and Adaptive low-dimensional subspace estimation (Alice). Experiments on LLaMA pre-training (up to 1B parameters) validate the effectiveness, showing faster and better convergence than existing memory-efficient baselines and Adam with little memory overhead. Notably, Alice achieves better than 2x faster convergence over Adam, while RACS delivers strong performance on the 1B model with SGD-like memory.

Training large language models (LLMs) typically relies on adaptive optimizers like Adam (Kingma & Ba, 2015) which store additional state information to accelerate convergence but incur significant memory overhead. Recent efforts, such as SWAN (Ma et al., 2024) address this by eliminating the need for optimizer states while achieving performance comparable to Adam via a multi-step preprocessing procedure applied to instantaneous gradients. Motivated by the success of SWAN, we introduce a novel framework for designing stateless optimizers that normalizes stochastic gradients according to multiple norms. To achieve this, we propose a simple alternating scheme to enforce the normalization of gradients w.r.t these norms. We show that our procedure can produce, up to an arbitrary precision, a fixed-point of the problem, and that SWAN is a particular instance of our approach with carefully chosen norms, providing a deeper understanding of its design. However, SWAN's computationally expensive whitening/orthogonalization step limit its practicality for large LMs. Using our principled perspective, we develop of a more efficient, scalable, and practical stateless optimizer. Our algorithm relaxes the properties of SWAN, significantly reducing its computational cost while retaining its memory efficiency, making it applicable to training large-scale models. Experiments on pre-training LLaMA models with up to 1 billion parameters demonstrate a 3X speedup over Adam with significantly reduced memory requirements, outperforming other memory-efficient baselines.

Recent years have seen an increase in the use of gigapixel-level image and video capture systems and benchmarks with high-resolution wide (HRW) shots. However, unlike close-up shots in the MS COCO dataset, the higher resolution and wider field of view raise unique challenges, such as extreme sparsity and huge scale changes, causing existing close-up detectors inaccuracy and inefficiency. In this paper, we present a novel model-agnostic sparse vision transformer, dubbed SparseFormer, to bridge the gap of object detection between close-up and HRW shots. The proposed SparseFormer selectively uses attentive tokens to scrutinize the sparsely distributed windows that may contain objects. In this way, it can jointly explore global and local attention by fusing coarse- and fine-grained features to handle huge scale changes. SparseFormer also benefits from a novel Cross-slice non-maximum suppression (C-NMS) algorithm to precisely localize objects from noisy windows and a simple yet effective multi-scale strategy to improve accuracy. Extensive experiments on two HRW benchmarks, PANDA and DOTA-v1.0, demonstrate that the proposed SparseFormer significantly improves detection accuracy (up to 5.8%) and speed (up to 3x) over the state-of-the-art approaches.

Adaptive optimizers such as Adam (Kingma & Ba, 2015) have been central to the success of large language models. However, they often require to maintain optimizer states throughout training, which can result in memory requirements several times greater than the model footprint. This overhead imposes constraints on scalability and computational efficiency. Stochastic Gradient Descent (SGD), in contrast, is a stateless optimizer, as it does not track state variables during training. Consequently, it achieves optimal memory efficiency. However, its capability in LLM training is limited (Zhao et al., 2024b). In this work, we show that pre-processing SGD in a stateless manner can achieve the same performance as the Adam optimizer for LLM training, while drastically reducing the memory cost. Specifically, we propose to pre-process the instantaneous stochastic gradients using normalization and whitening. We show that normalization stabilizes gradient distributions, and whitening counteracts the local curvature of the loss landscape. This results in SWAN (SGD with Whitening And Normalization), a stochastic optimizer that eliminates the need to store any optimizer states. Empirically, SWAN has the same memory footprint as SGD, achieving $\approx 50\%$ reduction on total end-to-end memory compared to Adam. In language modeling tasks, SWAN demonstrates comparable or even better performance than Adam: when pre-training the LLaMA model with 350M and 1.3B parameters, SWAN achieves a 2x speedup by reaching the same evaluation perplexity using half as many tokens.

Recent advancements in large language models (LLMs) have significantly advanced text-to-SQL systems. However, most LLM-based methods often narrowly focus on SQL generation, neglecting the complexities of real-world conversational queries. This oversight can lead to unreliable responses, particularly for ambiguous questions that cannot be directly addressed with SQL. To bridge this gap, we propose MMSQL, a comprehensive test suite designed to evaluate the question classification and SQL generation capabilities of LLMs by simulating real-world scenarios with diverse question types and multi-turn Q\&A interactions. Using MMSQL, we assessed the performance of popular LLMs, including both open-source and closed-source models, and identified key factors impacting their performance in such scenarios. Moreover, we introduce an LLM-based multi-agent framework that employs specialized agents to identify question types and determine appropriate answering strategies. Our experiments demonstrate that this approach significantly enhances the model's ability to navigate the complexities of conversational dynamics, effectively handling the diverse and complex nature of user queries.

Long-term time-series forecasting is essential for planning and decision-making in economics, energy, and transportation, where long foresight is required. To obtain such long foresight, models must be both efficient and effective in processing long sequence. Recent advancements have enhanced the efficiency of these models; however, the challenge of effectively leveraging longer sequences persists. This is primarily due to the tendency of these models to overfit when presented with extended inputs, necessitating the use of shorter input lengths to maintain tolerable error margins. In this work, we investigate the multiscale modeling method and propose the Logsparse Decomposable Multiscaling (LDM) framework for the efficient and effective processing of long sequences. We demonstrate that by decoupling patterns at different scales in time series, we can enhance predictability by reducing non-stationarity, improve efficiency through a compact long input representation, and simplify the architecture by providing clear task assignments. Experimental results demonstrate that LDM not only outperforms all baselines in long-term forecasting benchmarks, but also reducing both training time and memory costs.

Tracking multiple tiny objects is highly challenging due to their weak appearance and limited features. Existing multi-object tracking algorithms generally focus on single-modality scenes, and overlook the complementary characteristics of tiny objects captured by multiple remote sensors. To enhance tracking performance by integrating complementary information from multiple sources, we propose a novel framework called {HGT-Track (Heterogeneous Graph Transformer based Multi-Tiny-Object Tracking)}. Specifically, we first employ a Transformer-based encoder to embed images from different modalities. Subsequently, we utilize Heterogeneous Graph Transformer to aggregate spatial and temporal information from multiple modalities to generate detection and tracking features. Additionally, we introduce a target re-detection module (ReDet) to ensure tracklet continuity by maintaining consistency across different modalities. Furthermore, this paper introduces the first benchmark VT-Tiny-MOT (Visible-Thermal Tiny Multi-Object Tracking) for RGB-T fused multiple tiny object tracking. Extensive experiments are conducted on VT-Tiny-MOT, and the results have demonstrated the effectiveness of our method. Compared to other state-of-the-art methods, our method achieves better performance in terms of MOTA (Multiple-Object Tracking Accuracy) and ID-F1 score. The code and dataset will be made available at https://github.com/xuqingyu26/HGTMT.

Large-scale generative models have achieved remarkable success in a number of domains. However, for sequential decision-making problems, such as robotics, action-labelled data is often scarce and therefore scaling-up foundation models for decision-making remains a challenge. A potential solution lies in leveraging widely-available unlabelled videos to train world models that simulate the consequences of actions. If the world model is accurate, it can be used to optimize decision-making in downstream tasks. Image-to-video diffusion models are already capable of generating highly realistic synthetic videos. However, these models are not action-conditioned, and the most powerful models are closedsource which means they cannot be finetuned. In this work, we propose to adapt pretrained video diffusion models to action-conditioned world models, without access to the parameters of the pretrained model. Our approach, AVID, trains an adapter on a small domain-specific dataset of action-labelled videos. AVID uses a learned mask to modify the intermediate outputs of the pretrained model and generate accurate action-conditioned videos. We evaluate AVID on video game and real-world robotics data, and show that it outperforms existing baselines for diffusion model adaptation. Our results demonstrate that if utilized correctly, pretrained video models have the potential to be powerful tools for embodied AI. Large generative models trained on web-scale data have driven rapid improvement in natural language processing (Brown, 2020; Touvron et al., 2023; Achiam et al., 2023), image generation (Rombach et al., 2022), and video generation (OpenAI, 2024).

In the industry, numerous tasks are deployed online. Traditional approaches often tackle each task separately by its own network, which leads to excessive costs for developing and scaling models, especially in the context of large language models. Although multi-task methods can save costs through parameter sharing, they often struggle to outperform single-task methods in real-world applications. To tackle these challenges, we present a three-stage multi-task learning framework for large language models. It involves task filtering, followed by fine-tuning on high-resource tasks, and finally fine-tuning on all tasks. We conducted comprehensive experiments in single-task and multi-task settings. Our approach, exemplified on different benchmarks, demonstrates that it is able to achieve performance comparable to the single-task method while reducing up to 90.9\% of its overhead.

While humans effortlessly discern intrinsic dynamics and adapt to new scenarios, modern AI systems often struggle. Current methods for visual grounding of dynamics either use pure neural-network-based simulators (black box), which may violate physical laws, or traditional physical simulators (white box), which rely on expert-defined equations that may not fully capture actual dynamics. We propose the Neural Material Adaptor (NeuMA), which integrates existing physical laws with learned corrections, facilitating accurate learning of actual dynamics while maintaining the generalizability and interpretability of physical priors. Additionally, we propose Particle-GS, a particle-driven 3D Gaussian Splatting variant that bridges simulation and observed images, allowing back-propagate image gradients to optimize the simulator. Comprehensive experiments on various dynamics in terms of grounded particle accuracy, dynamic rendering quality, and generalization ability demonstrate that NeuMA can accurately capture intrinsic dynamics.