Roadside vision centric 3D object detection has received increasing attention in recent years. It expands the perception range of autonomous vehicles, enhances the road safety. Previous methods focused on predicting per-pixel height rather than depth, making significant gains in roadside visual perception. While it is limited by the perspective property of near-large and far-small on image features, making it difficult for network to understand real dimension of objects in the 3D world. Bird’s Eye View (BEV) features and voxel features present the real distribution of objects in 3D world compared to the image features. However, BEV features tend to lose details due to the lack of explicit height information, and voxel features are computationally expensive. Inspired by this insight, an efficient framework learning height prediction in voxel features via transformer is proposed, dubbed HeightFormer. It groups the voxel features into local height sequences, and utilize attention mechanism to obtain height distribution prediction. Subsequently, the local height sequences are reassembled to generate accurate voxel features. The proposed method is applied to two large-scale roadside benchmarks, DAIR-V2X-I and Rope3D. Extensive experiments are performed and the HeightFormer outperforms the state-of-the-art methods in roadside vision centric 3D object detection task. Code will be athttps://github.com/zhangzhang2024/HeightFormer

Storage-compute disaggregation has become a mainstream paradigm in cloud computing, yet data lake workloads introduce distinct prefetching challenges: massive numbers of small files, interleaved multi-tenant streams, and frequent one-time accesses. Under these conditions, traditional sequential, correlation-based, and semantic prefetching methods become ineffective, leading to cache inefficiency and high latency. We propose SemanticPrefetcher, a lightweight, semantic-aware prefetching mechanism that incrementally constructs meaningful access streams at runtime. The system operates in three stages: it tokenizes file paths or object names into a unified semantic representation, clusters related requests into coherent streams despite multi-tenant interleaving, and detects naming regularities to predict future accesses. This design transforms mixed and seemingly disordered requests into predictable access flows without application modifications. Implemented on JuiceFS, SemanticPrefetcher reduces end-to-end execution time by up to 39.6% and read latency by 79.3% compared to state-of-the-art baselines. These results demonstrate that implicit semantics in file paths can be effectively leveraged for robust and efficient prefetching in cloud-scale data lakes.

Artificial Intelligence (AI) has demonstrated significant potential in healthcare, particularly in disease diagnosis and treatment planning. Recent progress in Medical Large Vision-Language Models (Med-LVLMs) has opened up new possibilities for interactive diagnostic tools. However, these models often suffer from factual hallucination, which can lead to incorrect diagnoses. Fine-tuning and retrieval-augmented generation (RAG) have emerged as methods to address these issues. However, the amount of high-quality data and distribution shifts between training data and deployment data limit the application of fine-tuning methods. Although RAG is lightweight and effective, existing RAG-based approaches are not sufficiently general to different medical domains and can potentially cause misalignment issues, both between modalities and between the model and the ground truth. In this paper, we propose a versatile multimodal RAG system, MMed-RAG, designed to enhance the factuality of Med-LVLMs. Our approach introduces a domain-aware retrieval mechanism, an adaptive retrieved contexts selection, and a provable RAG-based preference fine-tuning strategy. These innovations make the RAG process sufficiently general and reliable, significantly improving alignment when introducing retrieved contexts. Experimental results across five medical datasets (involving radiology, ophthalmology, pathology) on medical VQA and report generation demonstrate that MMed-RAG can achieve an average improvement of 43.8% in factual accuracy in the factual accuracy of Med-LVLMs.

Reinforcement Learning from Human Feedback (RLHF) has shown promise in aligning large language models (LLMs). Yet its reliance on a singular reward model often overlooks the diversity of human preferences. Recent approaches address this limitation by leveraging multi-dimensional feedback to fine-tune corresponding reward models and train LLMs using reinforcement learning. However, the process is costly and unstable, especially given the competing and heterogeneous nature of human preferences. In this paper, we propose Mixing Preference Optimization (MPO), a post-processing framework for aggregating single-objective policies as an alternative to both multi-objective RLHF (MORLHF) and MaxMin-RLHF. MPO avoids alignment from scratch. Instead, it log-linearly combines existing policies into a unified one with the weight of each policy computed via a batch stochastic mirror descent. Empirical results demonstrate that MPO achieves balanced performance across diverse preferences, outperforming or matching existing models with significantly reduced computational costs.

Collision-free planning is essential for bipedal robots operating within unstructured environments. This paper presents a real-time Model Predictive Control (MPC) frame-work that addresses both body and foot avoidance for dynamic bipedal robots. Our contribution is two-fold: we introduce (1) a novel formulation for adjusting step timing to facilitate faster body avoidance and (2) a novel 3D foot-avoidance formulation that implicitly selects swing trajectories and footholds that either steps over or navigate around obstacles with awareness of Center of Mass (COM) dynamics. We achieve body avoidance by applying a half-space relaxation of the safe region but introduce a switching heuristic based on tracking error to detect a need to change foot-timing schedules. To enable foot avoidance and viable landing footholds on all sides of foot-level obstacles, we decompose the non-convex safe region on the ground into several convex polygons and use Mixed-Integer Quadratic Programming to determine the optimal candidate. We found that introducing a soft minimum-travel-distance constraint is effective in preventing the MPC from being trapped in local minima that can stall half-space relaxation methods behind obstacles. We demonstrated the proposed algorithms on multibody simulations on the bipedal robot platforms, Cassie and Digit, as well as hardware experiments on Digit.

Providing timely and personalized guidance for students’ programming assignments, particularly by indicating fine-grained error locations with explanations, offers significant practical value for helping students complete assignments and enhance their learning outcomes. In recent years, various automated Fault Localization (FL) techniques, particularly those leveraging Large Language Models (LLMs), have demonstrated promising results in identifying errors in programs. However, existing fault localization techniques face challenges when applied to educational contexts. Most approaches operate at the method level without explanatory feedback, resulting in granularity too coarse for students who need actionable insights to identify and fix their errors. While some approaches attempt line-level fault localization, they often depend on predicting line numbers directly in numerical form, which is ill-suited to LLMs. To address these challenges, we propose FLAME, a fine-grained, explainable Fault Localization method tailored for programming assignments via LLM-guided Annotation and Model Ensemble. FLAME leverages rich contextual information specific to programming assignments to guide LLMs in identifying faulty code lines. Instead of directly predicting line numbers, we prompt the LLM to annotate faulty code lines with detailed explanations, enhancing both localization accuracy and educational value. To further improve reliability, we introduce a weighted multi-model voting strategy that aggregates results from multiple LLMs to determine the suspiciousness of each code line. Extensive experimental results demonstrate that FLAME outperforms state-of-the-art fault localization baselines on programming assignments, successfully localizing 207 more faults at top-1 over the best-performing baseline. Beyond educational contexts, FLAME also generalizes effectively to general-purpose software codebases, outperforming all baselines on the Defects4J benchmark.

Transformer models have become the dominant backbone for sequence modeling, leveraging self-attention to produce contextualized token representations. These are typically aggregated into fixed-size vectors via pooling operations for downstream tasks. While much of the literature has focused on attention mechanisms, the role of pooling remains underexplored despite its critical impact on model behavior. In this paper, we introduce a theoretical framework that rigorously characterizes the expressivity of Transformer-based models equipped with widely used pooling methods by deriving closed-form bounds on their representational capacity and the ability to distinguish similar inputs. Our analysis extends to different variations of attention formulations, demonstrating that these bounds hold across diverse architectural variants. We empirically evaluate pooling strategies across tasks requiring both global and local contextual understanding, spanning three major modalities: computer vision, natural language processing, and time-series analysis. Results reveal consistent trends in how pooling choices affect accuracy, sensitivity, and optimization behavior. Our findings unify theoretical and empirical perspectives, providing practical guidance for selecting or designing pooling mechanisms suited to specific tasks. This work positions pooling as a key architectural component in Transformer models and lays the foundation for more principled model design beyond attention alone.

Autonomous underwater vehicles (AUVs) have become an important technology in the field of the Internet of Underwater Things (IoUT). However, the complexity and unknown nature of the underwater environment poses a great challenge to the autonomous operation of AUVs. An efficient and stable path planning algorithm is the key for AUVs to achieve autonomous operation. To address the above problems, this paper proposes an Improved Exponential Distribution Optimizer (IEDO) for three-dimensional path planning. In the proposed algorithm, population initialization, the optimization algorithm itself and the local optimum problem, are all addressed and improved. The algorithm population is first initialized using an oriented initialization method that obeys a Gaussian distribution to improve the efficiency of the algorithm in the early stages. Second, for the exponential distribution optimizer algorithm itself, the convergence rate of the algorithm is further improved by adding the guided solution generated by its iterative process to the iterative selection of the population. Finally, the crossover-mutation idea of the genetic algorithm is integrated to improve the global search ability of the population and avoid falling into the local optimum problem. In terms of algorithm validation, two real seabed terrain datasets are used for a simulation verification of the algorithm, and compared with the existing algorithms. The results prove that the IEDO algorithm proposed in this paper has a strong convergence speed, strong global search capability and good path qualities.

Neural combinatorial optimization aims to use neural networks to speed up the solving process of combinatorial optimization problems, i.e., finding the optimal solution of a problem instance from a finite set of feasible solutions that minimize a given objective function. Recently, researchers have applied convolutional neural networks to predict the optimal solution's cost (defined by the objective function) to give as extra input to an exact solver to speed up the solving process. In this paper, we investigate whether graph representations that explicitly model the inherent constraints in combinatorial optimization problems would improve the performance of predicting the optimal solution's cost. Specifically, we use graph neural networks with neighborhood aggregation and graph Transformer models to capture and embed the knowledge in the graph representations of combinatorial optimization problems. We also propose a benchmark dataset containing the Traveling Salesman Problem (TSP) and Job-Shop Scheduling Problem (JSSP), and through the empirical evaluation, we show that graph Transformer models achieve an average loss decrease of 61.05% on TSP and 66.53% on JSSP compared to the baseline convolutional neural networks.

In real-world settings, bipedal robots must avoid collisions with people and their environment. Further, a biped can choose between modes of avoidance: (1) adjust its pose while standing or (2) step to gain maneuverability. We present a real-time motion planner and multibody control framework for dynamic bipedal robots that avoids multiple moving obstacles and automatically switches between standing and stepping modes as necessary. By leveraging a reduced-order model (i.e. Linear Inverted Pendulum Model) and a half-space relaxation of the safe region, the planner is formulated as a convex optimization problem (i.e. Quadratic Programming) that can be used for real-time application with Model-Predictive-Control (MPC). To facilitate mode switching, we introduce center-of-pressure related slack-variables to the convex planning optimization that both shapes the planning cost function and provides a mode switching criterion for dynamic locomotion. Finally, we implement the proposed algorithm on a 3D Cassie bipedal robot and present hardware experiments showing real-time bipedal standing avoidance, stepping avoidance, and automatic switching of avoidance modes.

Constructing unbiased estimators from Markov chain Monte Carlo (MCMC) outputs is a difficult problem that has recently received a lot of attention in the statistics and machine learning communities. However, the current unbiased MCMC framework only works when the quantity of interest is an expectation, which excludes many practical applications. In this paper, we propose a general method for constructing unbiased estimators for functions of expectations and extend it to construct unbiased estimators for nested expectations. Our approach combines and generalizes the unbiased MCMC and Multilevel Monte Carlo (MLMC) methods. In contrast to traditional sequential methods, our estimator can be implemented on parallel processors. We show that our estimator has a finite variance and computational complexity and can achieve $\varepsilon$-accuracy within the optimal $O(1/\varepsilon^2)$ computational cost under mild conditions. Numerical experiments confirm our theoretical findings and demonstrate the benefits of unbiased estimators in the massively parallel regime.

Transformer models have great potential in Graph Representation Learning (GRL) for efficiently scaling the learning process on large datasets and solving many challenges presented in Graph Neural Networks, e.g., oversmoothing and suspended animation. To represent each node of a graph, Transformer models as input usually take a node together with the node context, i.e., a set of other nodes that serve as learning context for the target node. However, current GRL Transformer models mainly consider the graph topology when selecting the node context for each target node. In this work, we demonstrate the important role of node features in selecting the node context. Specifically, we propose a hybrid approach for selecting node context that considers both the graph topology and the semantic similarities between node features. Through the empirical evaluations, we show the advantages of our hybrid node context selection method for a downstream classification task on various datasets compared to selection methods that only consider graph topology or semantic similarities. The best classification accuracy improvements of our proposed hybrid methods over the baseline methods on each dataset range from 0.77% to 6.05%.

We present a real-time motion planner that avoids multiple moving obstacles without knowing their dynamics or intentions. This method uses convex optimization to generate trajectories for linear plant models over a planning horizon (i.e. model-predictive control). While convex optimizations allow for fast planning, obstacle avoidance can be challenging to incorporate because Euclidean distance calculations tend to break convexity. By using a half-space convex relaxation, our planner reasons about an approximated distance-to-obstacle measure that is linear in its decision variables and preserves convexity. Further, by iteratively updating the relaxation over the planning horizon, the half-space approximation is improved, enabling nimble avoidance maneuvers. We further augment avoidance performance with a soft penalty slack-variable for-mulation that introduces a piecewise quadratic cost. As a proof of concept, we demonstrate the planner on double-integrator models in both single-agent and multi-agent tasks-avoiding multiple obstacles and other agents in 2D and 3D environments. We show extensions to legged locomotion by bipedally walking around obstacles in simulation using the Linear Inverted Pendulum Model (LIPM). We then present two sets of hardware experiments showing real-time obstacle avoid-ance with quadcopter drones: (1) avoiding a 10m/s swinging pendulum and (2) dodging a chasing drone.

Reinforcement Learning (RL) has shown promising results across various domains. However, applying it to develop gameplaying agents is challenging due to sparsity of extrinsic rewards, where agents get rewards from the environments only at the end of game levels. Previous works have shown that using intrinsic rewards is an effective way to deal with such cases. Intrinsic rewards allow to incorporate basic skills in agent policies to better generalize over various game levels. In a gameplay, it is common that certain actions (skills) are observed more often than others, which leads to a biased selection of actions. This problem boils down to a normalization issue in formulating the skill-based reward function. In this paper, we propose a novel solution to this problem by taking into account the frequency of all skills in the reward function. We show that our method improves the performance of agents by enabling them to select effective skills up to 2.5 times more frequently than that of the state-of-the-art in the context of the match-3 game Candy Crush Friends Saga.

The Job-Shop Scheduling Problem (JSSP) is a well-known optimization problem with plenty of existing solutions. Although remarkable progress has been made in addressing the problem, most of the solutions require input from human experts. Deep Learning techniques, on the other hand, have proven successful in acquiring knowledge from data without using step-by-step instructions from humans. In this work, we propose a novel solution, called ConvJSSP, by applying Deep Learning to speed up the solving process of JSSPs and to reduce the need for human involvement. In ConvJSSP, we train a Convolutional Neural Network model for predicting the optimal makespan of JSSPs, and use the predicted makespan to accelerate the JSSP solving schema. Through the experiments, we compare several JSSP solving methods based on ConvJSSP approach with a state-of-the-art solution as a baseline, and show that ConvJSSP speeds up the problem solving up to 9% compared to the baseline method.

In this paper, we for the first time construct three cubic rotation symmetric bent functions in 2k+3, k ≥ 0, variables. Our work solves the open problem left by Gao et al. (IEEE TIT 58(7): 4908–4913, 2012).