Mixture-of-Agents (MoA) improves LLM performance through layered collaboration, but its dense topology raises costs and latency. Existing methods employ LLM judges to filter responses, yet still require all models to perform inference before judging, failing to cut costs effectively. They also lack model selection criteria and struggle with large model pools, where full inference is costly and can exceed context limits. To address this, we propose RouteMoA, an efficient mixture-of-agents framework with dynamic routing. It employs a lightweight scorer to perform initial screening by predicting coarse-grained performance from the query, narrowing candidates to a high-potential subset without inference. A mixture of judges then refines these scores through lightweight self- and cross-assessment based on existing model outputs, providing posterior correction without additional inference. Finally, a model ranking mechanism selects models by balancing performance, cost, and latency. RouteMoA outperforms MoA across varying tasks and model pool sizes, reducing cost by 89.8% and latency by 63.6% in the large-scale model pool. Code is available at https://github.com/Jize-W/RouteMoA.

We propose a semantic-aware neural reconstruction method to generate 3D high-fidelity models from sparse images. To tackle the challenge of severe radiance ambiguity caused by mismatched features in sparse input, we enrich neural implicit representations by adding patch-based semantic logits that are optimized together with the signed distance field and the radiance field. A novel regularization based on the geometric primitive masks is introduced to mitigate shape ambiguity. The performance of our approach has been verified in experimental evaluation. The average chamfer distances of our reconstruction on the DTU dataset can be reduced by 44% for SparseNeuS and 20% for VolRecon. When working as a plugin for those dense reconstruction baselines such as NeuS and Neuralangelo, the average error on the DTU dataset can be reduced by 69% and 68% respectively.

Computer-aided design (CAD) significantly enhances the efficiency, accuracy, and innovation of design processes by enabling precise 2D and 3D modeling, extensive analysis, and optimization. Existing methods for creating CAD models rely on latent vectors or point clouds, which are difficult to obtain, and storage costs are substantial. Recent advances in Multimodal Large Language Models (MLLMs) have inspired researchers to use natural language instructions and images for CAD model construction. However, these models still struggle with inferring accurate 3D spatial location and orientation, leading to inaccuracies in determining the spatial 3D starting points and extrusion directions for constructing geometries. This work introduces CAD-GPT, a CAD synthesis method with spatial reasoning-enhanced MLLM that takes either a single image or a textual description as input. To achieve precise spatial inference, our approach introduces a 3D Modeling Spatial Mechanism. This method maps 3D spatial positions and 3D sketch plane rotation angles into a 1D linguistic feature space using a specialized spatial unfolding mechanism, while discretizing 2D sketch coordinates into an appropriate planar space to enable precise determination of spatial starting position, sketch orientation, and 2D sketch coordinate translations. Extensive experiments demonstrate that CAD-GPT consistently outperforms existing state-of-the-art methods in CAD model synthesis, both quantitatively and qualitatively.

Document Information Extraction (DIE) aims to extract structured information from Visually Rich Documents (VRDs). Previous full-training approaches have demonstrated strong performance but may struggle with generalization to unseen data. In contrast, training-free methods leverage powerful pre-trained models like Large Language Models (LLMs) to address various downstream tasks with only a few examples. Nonetheless, training-free methods for DIE encounter two primary challenges: (1) understanding the complex relationship between layout and textual elements in VRDs, and (2) providing accurate guidance to pre-trained models. To address these challenges, we propose SAmple-centric In-context Learning (SAIL). SAIL introduces a fine-grained entity-level textual similarity to facilitate in-depth text analysis by LLMs and incorporates layout similarity to enhance the analysis of layouts in VRDs. Moreover, SAIL formulates a unified In-Context Learning (ICL) prompt template for various sample-centric examples, enabling tailored prompts that deliver precise guidance to pre-trained models for each sample. Extensive experiments on FUNSD, CORD, and SROIE benchmarks with various base models (e.g., LLMs) indicate that our SAIL outperforms training-free baselines, even closer to the full-training methods, showing the superiority and generalization of our method.

Recent advancements, such as DeepSeek-Prover-V2-671B and Kimina-Prover-Preview-72B, demonstrate a prevailing trend in leveraging reinforcement learning (RL)-based large-scale training for automated theorem proving. Surprisingly, we discover that even without any training, careful neuro-symbolic coordination of existing off-the-shelf reasoning models and tactic step provers can achieve comparable performance. This paper introduces DSP+, an improved version of the Draft, Sketch, and Prove framework, featuring a fine-grained and integrated neuro-symbolic enhancement for each phase: (1) In the draft phase, we prompt reasoning models to generate concise natural-language subgoals to benefit the sketch phase, removing thinking tokens and references to human-written proofs; (2) In the sketch phase, subgoals are autoformalized with hypotheses to benefit the proving phase, and sketch lines containing syntactic errors are masked according to predefined rules; (3) In the proving phase, we tightly integrate symbolic search methods like Aesop with step provers to establish proofs for the sketch subgoals. Experimental results show that, without any additional model training or fine-tuning, DSP+ solves 80.7%, 32.8%, and 24 out of 644 problems from miniF2F, ProofNet, and PutnamBench, respectively, while requiring fewer budgets compared to state-of-the-arts. DSP+ proves imo_2019_p1, an IMO problem in miniF2F that is not solved by any prior work. Additionally, DSP+ generates proof patterns comprehensible by human experts, facilitating the identification of formalization errors; For example, eight wrongly formalized statements in miniF2F are discovered. Our results highlight the potential of classical reasoning patterns besides the RL-based training. All components will be open-sourced.

Recent years have witnessed significant advancements in industrial anomaly detection (IAD) thanks to existing anomaly detection datasets. However, the large performance gap between these benchmarks and real industrial practice reveals critical limitations in existing datasets. We argue that the mismatch between current datasets and real industrial scenarios becomes the primary barrier to practical IAD deployment. To this end, we propose ReinAD dataset, a comprehensive contrastive dataset towards Real-world industrial Anomaly Detection. Our dataset prioritizes three critical real-world requirements: 1) Contrast-based anomaly definition that is essential for industrial practice, 2) Fine-grained unaligned image pairs reflecting real inspections, and 3) Large-scale data from active production lines spanning multiple industrial categories. Based on our dataset, we introduce the ReinADNet. It takes both normal reference and test images as inputs, achieving anomaly detection through normal-anomaly comparison. To address the fine-grained and unaligned properties of real industrial scenes, our method integrates pyramidal similarity aggregation for comprehensive anomaly characterization and global-local feature fusion for spatial misalignment tolerance. Our method outperforms all baselines on the ReinAD dataset (e.g., 64.5% v.s. 59.5% in 1-shot image-level AP) under all settings. Extensive experiments across several datasets demonstrate our dataset's challenging nature and our method's superior generalization. This work provides a solid foundation for practical industrial anomaly detection. Dataset and code are available at https://tocmac.github.io/ReinAD.

In developing general-purpose agents, significant focus has been placed on integrating large language models (LLMs) with various tools. This poses a challenge to the tool-use capabilities of LLMs. However, there are evident gaps between existing tool evaluations and real-world scenarios. Current evaluations often use AI-generated queries, single-step tasks, dummy tools, and text-only inputs, which fail to reveal the agents' real-world problem-solving abilities effectively. To address this, we propose GTA, a benchmark for General Tool Agents, featuring three main aspects: (i) Real user queries: human-written queries with simple real-world objectives but implicit tool-use, requiring the LLM to reason the suitable tools and plan the solution steps. (ii) Real deployed tools: an evaluation platform equipped with tools across perception, operation, logic, and creativity categories to evaluate the agents' actual task execution performance. (iii) Real multimodal inputs: authentic image files, such as spatial scenes, web page screenshots, tables, code snippets, and printed/handwritten materials, used as the query contexts to align with real-world scenarios closely. We designed 229 real-world tasks and executable tool chains to evaluate mainstream LLMs. Our findings show that real-world user queries are challenging for existing LLMs, with GPT-4 completing less than 50\% of the tasks and most LLMs achieving below 25\%. This evaluation reveals the bottlenecks in the tool-use capabilities of current LLMs in real-world scenarios, which is beneficial for the advancement of general-purpose tool agents. Dataset and code are available at https://github.com/open-compass/GTA.

Since the last decade, deep neural networks have shown remarkable capability in learning representations. The recently proposed neural ordinary differential equations (NODEs) can be viewed as the continuous-time equivalence of residual neural networks. It has been shown that NODEs have a tremendous advantage over the conventional counterparts in terms of spatial complexity for modeling continuous-time processes. However, existing NODEs methods entail their final time to be specified in advance, precluding the models from choosing a desirable final time and limiting their expressive capabilities. In this article, we propose learnable final-time (LFT) NODEs to overcome this limitation. LFT rebuilds the NODEs learning process as a final-time-free optimal control problem and employs the calculus of variations to derive the learning algorithm of NODEs. In contrast to existing NODEs methods, the new approach empowers the NODEs models to choose their suitable final time, thus being more flexible in adjusting the model depth for given tasks. Additionally, we analyze the gradient estimation errors caused by numerical ordinary differential equations (ODEs) solvers and employ checkpoint-based methods to obtain accurate gradients. We demonstrate the effectiveness of the proposed method with experimental results on continuous normalizing flows (CNFs) and feedforward models.

This work studies the problem of few-shot object counting, which counts the number of exemplar objects (i.e., described by one or several support images) occurring in the query image. The major challenge lies in that the target objects can be densely packed in the query image, making it hard to recognize every single one. To tackle the obstacle, we propose a novel learning block, equipped with a similarity comparison module and a feature enhancement module. Concretely, given a support image and a query image, we first derive a score map by comparing their projected features at every spatial position. The score maps regarding all support images are collected together and normalized across both the exemplar dimension and the spatial dimensions, producing a reliable similarity map. We then enhance the query feature with the support features by employing the developed point-wise similarities as the weighting coefficients. Such a design encourages the model to inspect the query image by focusing more on the regions akin to the support images, leading to much clearer boundaries between different objects. Extensive experiments on various benchmarks and training setups suggest that we surpass the state-of-the-art methods by a sufficiently large margin. For instance, on a recent large-scale FSC-147 dataset, we surpass the state-of-the-art method by improving the mean absolute error from 22.08 to 14.32 (35%↑). Code has been released in https://github.com/zhiyuanyou/SAFECount.

The crux of semi-supervised semantic segmentation is to assign adequate pseudo-labels to the pixels of unlabeled images. A common practice is to select the highly confident predictions as the pseudo ground-truth, but it leads to a problem that most pixels may be left unused due to their unreliability. We argue that every pixel matters to the model training, even its prediction is ambiguous. Intuitively, an unreliable prediction may get confused among the top classes (i.e., those with the highest probabilities), however, it should be confident about the pixel not belonging to the remaining classes. Hence, such a pixel can be convincingly treated as a negative sample to those most unlikely categories. Based on this insight, we develop an effective pipeline to make sufficient use of unlabeled data. Concretely, we separate reliable and unreliable pixels via the entropy of predictions, push each unreliable pixel to a category-wise queue that consists of negative samples, and manage to train the model with all candidate pixels. Considering the training evolution, where the prediction becomes more and more accurate, we adaptively adjust the threshold for the reliable-unreliable partition. Experimental results on various benchmarks and training settings demonstrate the superiority of our approach over the state-of-the-art alternatives.11Project: https://haochen-wang409.github.io/U2PL.

Despite the rapid advance of unsupervised anomaly detection, existing methods require to train separate models for different objects. In this work, we present UniAD that accomplishes anomaly detection for multiple classes with a unified framework. Under such a challenging setting, popular reconstruction networks may fall into an "identical shortcut", where both normal and anomalous samples can be well recovered, and hence fail to spot outliers. To tackle this obstacle, we make three improvements. First, we revisit the formulations of fully-connected layer, convolutional layer, as well as attention layer, and confirm the important role of query embedding (i.e., within attention layer) in preventing the network from learning the shortcut. We therefore come up with a layer-wise query decoder to help model the multi-class distribution. Second, we employ a neighbor masked attention module to further avoid the information leak from the input feature to the reconstructed output feature. Third, we propose a feature jittering strategy that urges the model to recover the correct message even with noisy inputs. We evaluate our algorithm on MVTec-AD and CIFAR-10 datasets, where we surpass the state-of-the-art alternatives by a sufficiently large margin. For example, when learning a unified model for 15 categories in MVTec-AD, we surpass the second competitor on the tasks of both anomaly detection (from 88.1% to 96.5%) and anomaly localization (from 89.5% to 96.8%). Code is available at https://github.com/zhiyuanyou/UniAD.

Distributed algorithms are gaining increasing research interests in the area of power system optimization and dispatch. Existing distributed power dispatch algorithms (DPDAs) usually assume that suppliers/consumers bid truthfully. However, this article shows the need for DPDAs to consider strategic players and to take account of their behavior deviation from what the DPDAs expect. To address this, we propose a distributed strategy update algorithm (DSUA) on top of a DPDA. The DSUA considers strategic suppliers who optimize their bids in a DPDA, using only the information accessible from a DPDA, that is, price. The DSUA also considers the cases when suppliers update bids alternately or simultaneously. Under both cases, we show the closeness of supplier bids to the Nash equilibrium via game-theoretic analysis as well as simulation.

In this article, a discrete-time distributed optimization algorithm is proposed for solving the economic dispatch (ED) problem with some groups of generator units to communicate over a connected graph, which is independent of the power system. The ED problem is converted to a distributed optimization problem with an objective of the sum of individual convex functions and constraints of local generators. Based on the optimal conditions, a class of distributed algorithms is designed to find the solution to the ED problem. The distributed algorithm can be realized as a multiagent system with a connected graph, whose convergence can be proved using the dynamic analysis method. Moreover, experiments with simulations are presented to demonstrate the performance of the proposed algorithm.

In this paper, we propose a Point Fractal Network (PF-Net), a novel learning-based approach for precise and high-fidelity point cloud completion. Unlike existing point cloud completion networks, which generate the overall shape of the point cloud from the incomplete point cloud and always change existing points and encounter noise and geometrical loss, PF-Net preserves the spatial arrangements of the incomplete point cloud and can figure out the detailed geometrical structure of the missing region(s) in the prediction. To succeed at this task, PF-Net estimates the missing point cloud hierarchically by utilizing a feature-points-based multi-scale generating network. Further, we add up multi-stage completion loss and adversarial loss to generate more realistic missing region(s). The adversarial loss can better tackle multiple modes in the prediction. Our experiments demonstrate the effectiveness of our method for several challenging point cloud completion tasks.

3D shape feature learning plays a pivotal role in both industry and academia. PointCNN is one of excellent neural networks for 3D object databases classification. Instead of selecting representative points arbitrarily in PointCNN, clustering-enhanced PointCNN proposed in this paper can make representative points more logical and efficient for point cloud classification learning. The proposed clustering-based selection approach is able to distinguish more features and catch more details from 3D shapes. Both K-Means and Gaussian-Mixture-Model (GMM) clustering methods are applied during the point selection period. Both methods have been tested on several public data sets, which substantiates the superior classification accuracy with comparable training time.

A neurodynamics-based algorithm is developed in this paper for solving multicoupled distributed optimization problems. In this formulation, each agent solves a local optimization problem with regard to its own cost function. The objective function is a sum of local convex subproblems, which are not necessarily strict convex and smooth. With communication between neighbors only, decision variables lie in a distributed manner and succeed to converge to the global optimum. The proposed method is suitable for solving large-scale problems in energy Internet management thanks to neural networks' capability of real-time computation. To verify the algorithm, a power-heat-gas multienergy system model is constructed. The simulation results on a 5-energy-hub multienergy system are demonstrated to show the efficacy of the algorithm.

In this paper, a distributed neurodynamic approach is proposed for constrained convex optimization. The objective function is a sum of local convex subproblems, whereas the constraints of these subproblems are coupled. Each local objective function is minimized individually with the proposed neurodynamic optimization approach. Through information exchange between connected neighbors only, all nodes can reach consensus on the Lagrange multipliers of all global equality and inequality constraints, and the decision variables converge to the global optimum in a distributed manner. Simulation results of two power system cases are discussed to substantiate the effectiveness and characteristics of the proposed approach.

This paper presents a two-time-scale neurodynamic approach to constrained minimax optimization using two coupled neural networks. One of the recurrent neural networks is used for minimizing the objective function and another is used for maximization. It is shown that the coupled neurodynamic systems operating in two different time scales work well for minimax optimization. The effectiveness and characteristics of the proposed approach are illustrated using several examples. Furthermore, the proposed approach is applied for H∞model predictive control.

This paper presents a neurodynamic optimization approach to bilevel quadratic programming (BQP). Based on the Karush-Kuhn-Tucker (KKT) theorem, the BQP problem is reduced to a one-level mathematical program subject to complementarity constraints (MPCC). It is proved that the global solution of the MPCC is the minimal one of the optimal solutions to multiple convex optimization subproblems. A recurrent neural network is developed for solving these convex optimization subproblems. From any initial state, the state of the proposed neural network is convergent to an equilibrium point of the neural network, which is just the optimal solution of the convex optimization subproblem. Compared with existing recurrent neural networks for BQP, the proposed neural network is guaranteed for delivering the exact optimal solutions to any convex BQP problems. Moreover, it is proved that the proposed neural network for bilevel linear programming is convergent to an equilibrium point in finite time. Finally, three numerical examples are elaborated to substantiate the efficacy of the proposed approach.

Abstract We present a method to design the deformation behavior of 3D printed models by an interactive tool, where the variation of bending elasticity at different regions of a model is realized by a change in shell thickness. Given a soft material to be used in 3D printing, we propose an experimental setup to acquire the bending behavior of this material on tubes with different diameters and thicknesses. The relationship between shell thickness and bending elasticity is stored in an echo state network using the acquired dataset. With the help of the network, an interactive design tool is developed to generate non‐uniformly hollowed models to achieve desired bending behaviors. The effectiveness of this method is verified on models fabricated by different 3D printers by studying whether their physical deformation can match the designed target shape.

This paper presents a neurodynamic optimization approach with two coupled recurrent neural networks for the synthesis of linear systems with fault detection via robust pole assignment. The proposed approach is shown to be capable of synthesizing control systems with robust state estimators and fault detection with parameter perturbation. The operating characteristics of the recurrent neural networks for state estimation and fault detection are demonstrated by using an illustrative example.

This paper presents new results on neurodynamic optimization approach to solve bilevel linear programming problems (BLPPs) with linear inequality constraints. A sub-gradient recurrent neural network is proposed for solving the BLPPs. It is proved that the state convergence time period is finite and can be quantitatively estimated. Compared with existing recurrent neural networks for BLPPs, the proposed neural network does not have any design parameter and can solve the BLPPs in finite time. Some numerical examples are introduced to show the effectiveness of the proposed neural network.

In this paper, a neurodynamic optimization approach is proposed for synthesizing high-order descriptor linear systems with state feedback control via robust pole assignment. With a new robustness measure serving as the objective function, the robust eigenstructure assignment problem is formulated as a pseudoconvex optimization problem. A neurodynamic optimization approach is applied and shown to be capable of maximizing the robust stability margin for high-order singular systems with guaranteed optimality and exact pole assignment. Two numerical examples and vehicle vibration control application are discussed to substantiate the efficacy of the proposed approach.

This paper introduces a perceptual model for determining 3D printing orientations. Additive manufacturing methods involving low-cost 3D printers often require robust branching support structures to prevent material collapse at overhangs. Although the designed shape can successfully be made by adding supports, residual material remains at the contact points after the supports have been removed, resulting in unsightly surface artifacts. Moreover, fine surface details on the fabricated model can easily be damaged while removing supports. To prevent the visual impact of these artifacts, we present a method to find printing directions that avoid placing supports in perceptually significant regions. Our model for preference in 3D printing direction is formulated as a combination of metrics including area of support, visual saliency, preferred viewpoint and smoothness preservation. We develop a training-and-learning methodology to obtain a closed-form solution for our perceptual model and perform a large-scale study. We demonstrate the performance of this perceptual model on both natural and man-made objects.

In this paper, a neurodynamic optimization approach is proposed for robust eigenstructure assignment problem of second-order descriptor systems via state feedback control. With a novel robustness measure serving as the objective function, the robust eigenstructure assignment problem is formulated as a pseudoconvex optimization problem. Two coupled recurrent neural networks are applied for solving the optimization problem with guaranteed optimality and exact pole assignment. Simulation results are included to substantiate the effectiveness of the proposed approach.

A neurodynamic optimization approach is proposed for robust pole assignment problem of second-order control systems via output feedback. With a suitable robustness measure serving as the objective function, the robust pole assignment problem is formulated as a quasi-convex optimization problem with linear constraints. Next, the problem further is reformulated as a convex feasibility problem. Two coupled recurrent neural networks are applied for solving the optimization problem with guaranteed optimality and exact pole assignment. Simulation results are included to substantiate the effectiveness of the proposed approach.

This paper presents a neurodynamic optimization approach to robust pole assignment for synthesizing linear control systems via state and output feedback. The problem is formulated as a pseudoconvex optimization problem with robustness measure: i.e., the spectral condition number as the objective function and linear matrix equality constraints for exact pole assignment. Two coupled recurrent neural networks are applied for solving the formulated problem in real time. In contrast to existing approaches, the exponential convergence of the proposed neurodynamics to global optimal solutions can be guaranteed even with lower model complexity in terms of the number of variables. Simulation results of the proposed neurodynamic approach for 11 benchmark problems are reported to demonstrate its superiority.

This paper presents new results on neurodynamic optimization approaches to robust pole assignment based on four alternative robustness measures. One or two recurrent neural networks are utilized to optimize these measures while making exact pole assignment. Compared with existing approaches, the present neurodynamic approaches can result in optimal robustness in most cases with one of the robustness measures. Simulation results of the proposed approaches for many benchmark problems are reported to demonstrate their performances.