Large Language Models (LLMs) extended to multi-modal inputs have led to Multi-modal LLMs (MLLMs) that perform strongly on vision-language tasks. Recent MLLMs adopt multi-resolution inputs to capture both global context and local details, but this substantially increases visual tokens and computational cost. Existing pruning methods reduce redundancy but are designed for single-resolution settings, overlooking the characteristics of multi-resolution tokens. We observe two key properties: tokens from different resolutions follow distinct distributions of information content, and tokens across resolutions exhibit mutual complementarity, such that pruning one type can often be compensated by the other. Based on this observation, we propose Multi-Resolution Token Pruning method (MR-Pruner), a training-free, graph-based pruning framework for multi-resolution MLLMs. MR-Pruner incorporates three components—Intra-resolution, Cross-resolution Token Scoring, and Informativeness-aware Token Pruning—that adaptively allocate pruning ratios and facilitate information propagation across resolutions. Experiments on eight benchmarks show that MR-Pruner achieves superior efficiency–performance trade-offs. For example, when only 10% of the visual tokens are retained, it leads to an average performance degradation of 3.6%. For reproducibility, the source code is available at https://github.com/gooriiie/MR-Pruner.

Modern neural network models commonly have residual connections, because they are helpful to achieve better performance. Due to their unconditional popularity, modifying them to achieve a better efficiency-accuracy trade-off is rarely studied in the literature. Motivated by this, we study how to get an efficient sub-network by rewiring a neural block having residual connections based on their inference paths. Based on this, we devise a new simulated annealing-based neural network rewiring method. Then, we construct a simple yet effective compression pipeline by combining this rewiring method and a recent channel pruning method. To demonstrate the effectiveness of the pipeline, we use object detection as the target task and consider YOLOv8 as the target model. We conduct experiments with two well-known datasets: VisDrone and PASCAL VOC. The results of the experiments demonstrate that our pipeline successfully outperforms the pruning method alone in most cases. Compared to YOLOv8 series, our method can offer more accurate models for VisDrone.

To leverage the complex structures within heterogeneous graphs, recent studies on heterogeneous graph embedding use a hyperbolic space, characterized by a constant negative curvature and exponentially increasing space, which aligns with the structural properties of heterogeneous graphs. However, despite heterogeneous graphs inherently possessing diverse power-law structures, most hyperbolic heterogeneous graph embedding models use a single hyperbolic space for the entire heterogeneous graph, which may not effectively capture the diverse power-law structures within the heterogeneous graph. To address this limitation, we propose Multi-hyperbolic Space-based heterogeneous Graph Attention Network (MSGAT), which uses multiple hyperbolic spaces to effectively capture diverse power-law structures within heterogeneous graphs. We conduct comprehensive experiments to evaluate the effectiveness of MSGAT. The experimental results demonstrate that MSGAT outperforms state-of-the-art baselines in various graph machine learning tasks, effectively capturing the complex structures of heterogeneous graphs.

As deep learning models become popular, there is a lot of need for deploying them to diverse device environments. Because it is costly to develop and optimize a neural network for every single environment, there is a line of research to search neural networks for multiple target environments efficiently. However, existing works for such a situation still suffer from requiring many GPUs and expensive costs. Motivated by this, we propose a novel neural network optimization framework named Bespoke for low-cost deployment. Our framework searches for a lightweight model by replacing parts of an original model with randomly selected alternatives, each of which comes from a pretrained neural network or the original model. In the practical sense, Bespoke has two significant merits. One is that it requires near zero cost for designing the search space of neural networks. The other merit is that it exploits the sub-networks of public pretrained neural networks, so the total cost is minimal compared to the existing works. We conduct experiments exploring Bespoke's the merits, and the results show that it finds efficient models for multiple targets with meager cost.

A distance sensitivity oracle is a data structure answering queries that ask the shortest distance from a node to another in a network expecting node/edge failures. It has been mainly studied in theory literature, but all the existing oracles for a directed graph suffer from prohibitive preprocessing time and space. Motivated by this, we develop two practical distance sensitivity oracles for directed graphs as variants of Transit Node Routing. The first oracle consists of a novel fault-tolerant index structure, which is used to construct a solution path and to detect and localize the impact of network failures, and an efficient query algorithm for it. The second oracle is made by applying the A* heuristics to the first oracle, which exploits lower bound distances to effectively reduce search space. In addition, we propose additional speed-up techniques to make our oracles faster with a slight loss of accuracy. We conduct extensive experiments with real-life datasets, which demonstrate that our oracles greatly outperform all of competitors in most cases. To the best of our knowledge, our oracles are the first distance sensitivity oracles that handle real-world graph data with million-level nodes.

A distance sensitivity oracle is a data structure answering queries that ask the shortest distance from a node to another in a network expecting node/edge failures. It has been mainly studied in theory literature, but all the existing oracles for a directed graph suffer from prohibitive preprocessing time and space. Motivated by this, we develop two practical distance sensitivity oracles for directed graphs as variants of Transit Node Routing, and effective speed-up techniques with a slight loss of accuracy. Extensive experiments demonstrate that our oracles greatly outperform all of competitors in most cases. To the best of our knowledge, our oracles are the first distance sensitivity oracles that handle real-world graph data with million-level nodes.

Influence maximization is introduced to maximize the profit of viral marketing in social networks. The weakness of influence maximization is that it does not distinguish specific users from others, even if some items can be only useful for the specific users. For such items, it is a better strategy to focus on maximizing the influence on the specific users. In this paper, we formulate an influence maximization problem as query processing to distinguish specific users from others. We show that the query processing problem is NP-hard and its objective function is submodular. We propose an expectation model for the value of the objective function and a fast greedy-based approximation method using the expectation model. For the expectation model, we investigate a relationship of paths between users. For the greedy method, we work out an efficient incremental updating of the marginal gain to our objective function. We conduct experiments to evaluate the proposed method with real-life datasets, and compare the results with those of existing methods that are adapted to the problem. From our experimental results, the proposed method is at least an order of magnitude faster than the existing methods in most cases while achieving high accuracy.

In the classic range aggregation problem, we have a set S of objects such that, given an interval I, a query counts how many objects of S are covered by I. Besides COUNT, the problem can also be defined with other aggregate functions, e.g., SUM, MIN, MAX and AVERAGE. This paper studies a novel variant of range aggregation, where an object can belong to multiple sets. A query (at runtime) picks any two sets, and aggregates on their intersection. More formally, let S1,...,Smbe m sets of objects. Given distinct set ids i, j and an interval I, a query reports how many objects in Si∩ Sjare covered by I. We call this problem range aggregation with set selection (RASS). Its hardness lies in that the pair (i, j) can have (2m) choices, rendering effective indexing a non-trivial task. 2 The RASS problem can also be defined with other aggregate functions, and generalized so that a query chooses more than 2 sets. We develop a system called RASS to power this type of queries. Our system has excellent efficiency in both theory and practice. Theoretically, it consumes linear space, and achieves nearly-optimal query time. Practically, it outperforms existing solutions on real datasets by a factor up to an order of magnitude. The paper also features a rigorous theoretical analysis on the hardness of the RASS problem, which reveals invaluable insight into its characteristics.