Spatio-temporal forecasting is fundamental to intelligent systems in transportation, climate science, and urban planning. However, training deep learning models on the massive, often redundant, datasets from these domains presents a significant computational bottleneck. Existing solutions typically focus on optimizing model architectures or optimizers, while overlooking the inherent inefficiency of the training data itself. This conventional approach of iterating over the entire static dataset each epoch wastes considerable resources on easy-to-learn or repetitive samples. In this paper, we explore a novel training-efficiency techniques, namely learning from complexity with dynamic sample pruning, ST-Prune, for spatio-temporal forecasting. Through dynamic sample pruning, we aim to intelligently identify the most informative samples based on the model's real-time learning state, thereby accelerating convergence and improving training efficiency. Extensive experiments conducted on real-world spatio-temporal datasets show that ST-Prune significantly accelerates the training speed while maintaining or even improving the model performance, and it also has scalability and universality.

Existing conversational memory systems rely on complex hierarchical summarization or reinforcement learning to manage long-term dialogue history, yet remain vulnerable to context dilution as conversations grow. In this work, we offer a different perspective: the primary bottleneck may lie not in memory architecture, but in the Signal Sparsity Effect within the latent knowledge manifold. Through controlled experiments, we identify two key phenomena: Decisive Evidence Sparsity, where relevant signals become increasingly isolated with longer sessions, leading to sharp degradation in aggregation-based methods; and Dual-Level Redundancy, where both inter-session interference and intra-session conversational filler introduce large amounts of non-informative content, hindering effective generation. Motivated by these insights, we propose , a minimalist framework that brings conversational memory back to basics, relying solely on retrieval and generation via Turn Isolation Retrieval (TIR) and Query-Driven Pruning (QDP). TIR replaces global aggregation with a max-activation strategy to capture turn-level signals, while QDP removes redundant sessions and conversational filler to construct a compact, high-density evidence set. Extensive experiments on multiple benchmarks demonstrate that achieves robust performance across diverse settings, consistently outperforming strong baselines while maintaining high efficiency in tokens and latency, establishing a new minimalist baseline for conversational memory.

The diversified route planning finds multiple paths that are sufficiently different from each other while as short as possible. It is of great significance to traffic alleviation through provided alternative routes during navigation. However, it is NP-H to find the optimal result and the existing solutions have either high quality (exact path enumeration-based) or high efficiency (alternative path), which is highly affected by the network properties and query parameters but has never been investigated before. Therefore, this paper proposes a hybrid diversified routing system that can handle any query efficiently with quality as high as possible. Specifically, we first analyze the path enumeration problem from the ground up and unify all the existing algorithms theoretically to identify the factors that affect algorithm performance. After that, we review and select the alternative path methods to identify the suitable ones for our system. Finally, we propose a query classification module to estimate the hardness of a query and determine how it should be processed. Extensive experiments on real-life networks validate the effectiveness and efficiency of our hybrid system compared with state-of-the-art solutions.

Trajectory computing, which encompasses trajectory data management and mining, has been widely studied owing to its role in practical applications such as location services, urban traffic, and public safety. Traditional methods, focusing on simplistic spatio-temporal features, face challenges of complex calculations, limited scalability, and inadequate adaptability to real-world complexities. In this paper, we present a comprehensive review of the development and recent advances in trajectory computing, from deep learning to the more recent large language models. We first define trajectory data and provide a brief overview of widely-used deep learning models. Systematically, we explore deep learning applications in trajectory management (pre-processing, storage, analysis, and visualization) and mining (trajectory-related forecasting, trajectory-related recommendation, trajectory classification, travel time estimation, anomaly detection, and mobility generation). Furthermore, we discuss emerging research directions and recent advancements in large models (represented by foundation models and large language models) for trajectory computing, which promise to reshape the next generation of trajectory computing. We also provide practical guidance on method selection for each task and analyze the cost–accuracy trade-offs of large models against specialized deep models. Additionally, we summarize application scenarios, public datasets, and toolkits. Finally, we outline current challenges in trajectory computing research and propose future directions. Relevant papers and open-source resources have been collated and are continuously

Trust relationships play a crucial role in various domains, such as social spam detection, retweet behavior analytics, and recommendation systems. Trust is often implicit and difficult to observe directly in the real world, as it is driven by people's underlying intentions and motivations. Therefore, when evaluating trust, it is critical to analyze not only user behavior data but also the intentions behind these behaviors that lead to trust. Existing trust evaluation methods often neglect the underlying reasons behind connections, such as shared hobbies or belonging to the same community. Therefore, these methods cannot differentiate the genuine intentions that lead to trust, resulting in an inaccurate evaluation of hidden trust relationships. To address this issue, we propose a novel Intent-based model for Trust Evaluation (INTRUST). This model can distinguish the intent behind high-order information in social communities using hypergraphs. Initially, we used hyperedges to represent high-order correlations between user-to-item and user-to-user interactions. Then, we construct K intent prototypes, which serve as foundational elements to build trust. Furthermore, we distinguish K-independent intent subgraphs from these high order correlations. To enhance the generalization and robustness of the model, we employ self-supervised learning and construct contrastive views at the node-level, hyperedge-level, and node hyperedge-level. Extensive experiments on real-world datasets demonstrate that our model outperforms state-of-the-art approaches in terms of trust evaluation accuracy and efficiency.

Reinforcement learning drives recent advances in LLM reasoning and agentic capabilities, yet current approaches struggle with both exploration and exploitation. Exploration suffers from low success rates on difficult tasks and high costs of repeated rollouts from scratch. Exploitation suffers from coarse credit assignment and training instability: Trajectory-level rewards penalize valid prefixes for later errors, and failure-dominated groups overwhelm the few positive signals, leaving optimization without constructive direction. To this end, we propose R^3L, Reflect-then-Retry Reinforcement Learning with Language-Guided Exploration, Pivotal Credit, and Positive Amplification. To synthesize high-quality trajectories, R^3L shifts from stochastic sampling to active synthesis via reflect-then-retry, leveraging language feedback to diagnose errors, transform failed attempts into successful ones, and reduce rollout costs by restarting from identified failure points. With errors diagnosed and localized, Pivotal Credit Assignment updates only the diverging suffix where contrastive signals exist, excluding the shared prefix from gradient update. Since failures dominate on difficult tasks and reflect-then-retry produces off-policy data, risking training instability, Positive Amplification upweights successful trajectories to ensure positive signals guide the optimization process. Experiments on agentic and reasoning tasks demonstrate 5% to 52% relative improvements over baselines while maintaining training stability. Our code is released at https://github.com/shiweijiezero/R3L.

Given a social network G = (V, E), the unconstrained profit maximization problem aims to identify a subset S subset of V that maximizes the net profit, defined as the expected influence spread F(S) of set S minus the associated cost c(S), i.e., Gamma(S)-c(S). However, this problem presupposes an unlimited budget, which is often impractical in real scenarios. Motivated by this, we investigate the budgeted profit maximization (BPM) problem by adding a budget constraint. Unfortunately, addressing the BPM problem with a theoretical approximation guarantee remains relatively under-explored in the literature. In response, assuming F(S) is known for any S subset of V, we propose an algorithm that guarantees returning a set S degrees such that Gamma(S degrees)-c(S degrees) >= (1-1/e ) Gamma(S-& lowast;/2) - c(S-& lowast;/2), where S-& lowast; denotes an optimal solution for BPM. Then, we develop a practical solution, which uses the reverse reachable set (RR-set) technique for influence estimation, without assuming knowledge of Gamma(S), while still maintaining a strong approximation guarantee. Additionally, similar to existing RR-set-based solutions for influence cascade-related problems, our RR-set-based solution relies on generating a large number of random RR-sets to accurately estimate Gamma(S). However, the existing RR-set generation method suffers from high memory stall rates due to its irregular memory access patterns, leaving room for further efficiency improvement. Therefore, we propose a new RR-set generation method that utilizes batch execution and cache prefetching. When a memory access is required, instead of stalling while waiting for data, the CPU first issues an asynchronous prefetch request to load the target data into the cache, and then switches to processing the generation of other RR-sets within the same batch, effectively hiding memory access latency. This method can be seamlessly integrated into existing RR-set-based solutions to improve their efficiency. Finally, we conduct extensive experiments on real, large-scale datasets to demonstrate the effectiveness and efficiency of our proposed solutions.

In multi-vector retrieval, both queries and data are represented as sets of high-dimensional vectors, enabling finer-grained semantic matching and improving retrieval quality over single-vector approaches. However, its practical adoption is held back by the lack of effective indexing algorithms. Existing work, attempting to reuse standard single-vector indexes, often fails to preserve multi-vector semantics or remains slow. In this work, we present GEM, a native indexing framework for multi-vector representations. The core idea is to construct a proximity graph directly over vector sets, preserving their fine-grained semantics while enabling efficient navigation. First, GEM designs a set-level clustering scheme. It associates each vector set with only its most informative clusters, effectively reducing redundancy without hurting semantic coverage. Then, it builds local proximity graphs within clusters and bridges them into a globally navigable structure. To handle the non-metric nature of multi-vector similarity, GEM decouples the graph construction metric from the final relevance score and injects semantic shortcuts to guide efficient navigation toward relevant regions. At query time, GEM launches beam search from multiple entry points and prunes paths early using cluster cues. To further enhance efficiency, a quantized distance estimation technique is used for both indexing and search. Across in-domain, out-of-domain, and multi-modal benchmarks, GEM achieves up to 16x speedup over state-of-the-art methods while matching or improving accuracy.

Data provenance (the process of determining the origin and derivation of data outputs) has applications across multiple domains including explaining database query results and auditing scientific workflows. Despite decades of research, provenance tracing remains challenging due to its high computational cost and storage requirements. In streaming systems such as Apache Flink, fine-grained provenance graphs can grow super-linearly with data volume, posing significant scalability challenges. We define temporal attribution, a new lightweight form of provenance, appropriate for certain tasks, such as monitoring dependencies between system components over time quantitatively. Temporal attribution enables time-focused analysis that does not require fine-grained, tuple-level dependency meta-data. Inspired by volume-based provenance tracking in Temporal Interaction Networks (TINs), we demonstrate TINs' applicability in succinctly modeling quantified data exchanges between dataflow operators in stream data processing systems and in processing workflows, in general, over time. We classify data into discrete and liquid types, define five temporal provenance query types, and propose a state-based indexing approach. Our vision outlines research directions toward making this new form of temporal attribution a practical tool for large-scale dataflow analytics.

Retrieval-augment generation is a prevalent strategy to mitigate hallucinations of LLMs. The attributable RAG (RAGQ) generates quotes for its answers. The quotes indicate which input contexts support the RAG to derive the answers, enhancing the answer's verifiability and trustworthiness. However, existing RAGQs exhibit significant degradation when dealing with questions that require multi-hop reasoning and multi-modal understanding, suffering from over-citation, implicit entity identification failure, and poor generalization. In this paper, we propose a novel RAGQ framework, namely QDRAG. QDRAG breaks down the input question into atomic subquestions to identify the implicit entities. Then, the reranker prunes context distractors to eliminate the downstream over-citation. To facilitate query decomposition, we propose two zero-shot approaches: QD-C and QD-R, which guide the QD MLLM to decompose the question based on context knowledge and retrieval rewards, respectively. One interesting finding is that finetuning on the QD task shows better generalizability compared to directly finetuning on the downstream RAGQ task. Experiments on four multi-modal QA benchmarks demonstrate QDRAG's efficacy in grounding answers and generating faithful citations. The framework significantly outperforms all the baselines on both in-domain and out-of-domain tests, even surpassing Gemini-Pro.

Enumerating the top- $k$ simple shortest path (KSP) is a fundamental searching strategy for many path-related applications. However, the efficiency of current solutions is not acceptable, especially when the required $k$ is large, which becomes a bottleneck for many downstream tasks like Diversified KSP (DkSP) and Constrained Shortest Path. Besides, it seems infeasible to build a KSP index specifically when $k$ is unknown beforehand. To break through these barriers, in this paper, we first propose an index $\boldsymbol{i}$ KSP for top-k path enumeration. Then we dive into the path concatenation relationship, and propose novel pruning techniques to efficiently detect the loopful and repeated paths for KSP computation. Finally, the experiment on real-life road networks demonstrates the effectiveness and efficiency of our algorithm over the state-of-the-art, and the application in DkSP shows the practicality of our algorithm.

3D object grounding localizes referred objects in a 3D scene from natural language. Unified instance-centric 3D-LLMs aim to solve grounding together with dialog, QA, and captioning, yet many rely on a single pointer-style grounding decision that compresses a relational instruction into one selection. This is brittle for fine-grained queries where multiple same-class candidates must be ruled out by context objects and spatial relations. We propose Structured Spatial Reasoning 3D-LLM (SSR3D-LLM), a structured grounding interface for unified 3D-LLMs. Given fixed Mask3D object proposals, the LLM writes a sequence of latent spatial reasoning steps and memory tokens from the query, and a geometry-aware scorer reads these latent steps in order to refine candidate rankings step by step with step-length masking. The latent steps are learned from standard benchmark target supervision with auxiliary referential-cue supervision during training, while inference uses only the input query and Mask3D proposals. Across ReferIt3D, ScanRefer, and Multi3DRef, SSR3D-LLM achieves the strongest results among unified 3D-LLM baselines, with substantial gains over the single-pointer QPG baseline on fine-grained grounding and consistent improvements over prior unified 3D-LLMs, while preserving the default language-task route.

Language agents are increasingly deployed over accumulating multimodal information, yet existing benchmarks assume a human-human form with sparse visuals and straightforward content, evaluating neither reasoning over authentic multimodal file interaction nor the interpretation of concealed user information. We therefore introduce M^3Exam, a query-centric multimodal conversational memory benchmark built on realistic user-agent interaction, with multi-dimensional evaluation spanning cross-modal grounding and implicit information inference. Benchmarking MLLMs and memory systems reveals persistent gaps in cross-modal grounding, cross session reasoning, and the efficiency cost of accumulating multimodal context. We further propose M^3Proctor, a multimodal memory method that detects query modality bias and consumes raw visual sources only on demand, improving accuracy by 13

The integration of Knowledge Graphs (KGs) into the Retrieval Augmentation Generation (RAG) framework has attracted significant interest, with early studies showing promise in mitigating hallucinations and improving model accuracy. However, a systematic understanding and comparative analysis of the rapidly emerging KG-RAG methods are still lacking. This paper seeks to lay the foundation for systematically answering the questions of when and how to use KG-RAG by analyzing its performance in various application scenarios associated with different technical configurations. After outlining the mind map using the KG-RAG framework and summarizing its popular pipeline, we conduct a pilot empirical study of KG-RAG works to reimplement and evaluate 6 KG-RAG methods across 9 datasets in diverse domains and scenarios, analyzing the impact of 9 KG-RAG configurations in combination with 17 LLMs. Our results underscore the critical role of appropriate application conditions and optimal configurations of KG-RAG components. The full appendix, data, and methods used in our paper, along with our reimplementation, are available on https://github.com/XujieYuan/Understanding-KG-RAG .

Speculative decoding accelerates LLM inference by drafting a tree of candidate continuations and verifying it in one target forward. Existing drafters fall into two camps with opposite weaknesses. Autoregressive drafters such as EAGLE-3 preserve dependence along each draft path but call the drafter once per tree depth, making drafting a non-trivial share of per-iteration latency. Parallel drafters cut drafter calls by predicting multiple future positions in one forward, but each position is predicted without seeing the others, producing paths the verifier rejects. In this paper, we propose SpecBlock, a block-iterative drafter that combines path dependence with cheap drafting. Each drafter forward produces K dependent positions and we call this a block. The draft tree grows through repeated block expansions. Two mechanisms explicitly carry path dependence to keep later draft positions accurate. Within each block, a layer-wise shift carries the previous position's hidden state into every decoder layer. Across blocks, each new block can start from any position of the previous block, inheriting its hidden state to extend the path. To spend verifier budget where acceptance is likely, a co-trained rank head replaces the fixed top-k tree by allocating per-position branching during drafting. To avoid training the drafter on prefixes it never produces at inference, a valid-prefix mask drops the loss at later positions once an earlier one is wrong. Beyond static drafting, a cost-aware bandit at deployment uses free verifier feedback to update the drafter selectively, only when the expected throughput gain exceeds the update cost. Experiments show that SpecBlock improves mean speedup by 8-13

Existing disentanglement methods face two main problems: First, they focus excessively on local structural features from direct node interactions while overlooking the overall graph structure, which limits accuracy. Second, they depend solely on backpropagation signals from recommendation tasks and lack direct supervision, making them prone to bias and overfitting. To address the problems, we propose the Intent Propagation Contrastive Collaborative Filtering (IPCCF) algorithm. Specifically, we propose a double-helix message propagation framework for richer node semantics and interaction modeling. We also develop a structure-aware disentanglement method, applying contrastive learning to align node representations from structural and intent perspectives. This provides direct supervision for disentanglement, reduces bias, and improves robustness. Experiments on three real-world graph datasets confirm the model's superiority.

Cellular Trajectory Map-Matching (CTMM) aims to align cellular location sequences to road networks, which is a necessary preprocessing in location-based services on web platforms like Google Maps, including navigation and route optimization. Current approaches mainly rely on ID-based features and region-specific data to learn correlations between cell towers and roads, limiting their adaptability to unexplored areas. To enable high-accuracy CTMM without additional training in target regions, Zero-shot CTMM requires to extract not only region-adaptive features, but also sequential and location uncertainty to alleviate positioning errors in cellular data. In this paper, we propose a pixel-based trajectory calibration assistant for zero-shot CTMM, which takes advantage of transferable geospatial knowledge to calibrate pixelated trajectory, and then guide the path-finding process at the road network level. To enhance knowledge sharing across similar regions, a Gaussian mixture model is incorporated into VAE, enabling the identification of scenario-adaptive experts through soft clustering. To mitigate high positioning errors, a spatial-temporal awareness module is designed to capture sequential features and location uncertainty, thereby facilitating the inference of approximate user positions. Finally, a constrained path-finding algorithm is employed to reconstruct the road ID sequence, ensuring topological validity within the road network. This process is guided by the calibrated trajectory while optimizing for the shortest feasible path, thus minimizing unnecessary detours. Extensive experiments demonstrate that our model outperforms existing methods in zero-shot CTMM by 16.8%.

Recent advances in Retrieval-Augmented Generation (RAG) have shifted from simple vector similarity to structure-aware approaches like HippoRAG, which leverage Knowledge Graphs (KGs) and Personalized PageRank (PPR) to capture multi-hop dependencies. However, these methods suffer from a "Static Graph Fallacy": they rely on fixed transition probabilities determined during indexing. This rigidity ignores the query-dependent nature of edge relevance, causing semantic drift where random walks are diverted into high-degree "hub" nodes before reaching critical downstream evidence. Consequently, models often achieve high partial recall but fail to retrieve the complete evidence chain required for multi-hop queries. To address this, we propose CatRAG, Context-Aware Traversal for robust RAG, a framework that builds on the HippoRAG 2 architecture and transforms the static KG into a query-adaptive navigation structure. We introduce a multi-faceted framework to steer the random walk: (1) Symbolic Anchoring, which injects weak entity constraints to regularize the random walk; (2) Query-Aware Dynamic Edge Weighting, which dynamically modulates graph structure, to prune irrelevant paths while amplifying those aligned with the query's intent; and (3) Key-Fact Passage Weight Enhancement, a cost-efficient bias that structurally anchors the random walk to likely evidence. Experiments across four multi-hop benchmarks demonstrate that CatRAG consistently outperforms state of the art baselines. Our analysis reveals that while standard Recall metrics show modest gains, CatRAG achieves substantial improvements in reasoning completeness, the capacity to recover the entire evidence path without gaps. These results reveal that our approach effectively bridges the gap between retrieving partial context and enabling fully grounded reasoning. Resources are available at https://github.com/kwunhang/CatRAG.

Spatio-temporal data proliferates in numerous real-world domains, such as transportation, weather, and energy. Spatio-temporal deep learning models aims to utilize useful patterns in such data to support tasks like prediction, imputation, and anomaly detection. However, previous one-to-one deep learning models designed for specific tasks typically require separate training for each use case, leading to increased computational and storage costs. To address this issue, one-to-many spatio-temporal foundation models have emerged, offering a unified framework capable of solving multiple spatio-temporal tasks. These foundation models achieve remarkable success by learning general knowledge with spatio-temporal data or transferring the general capabilities of pre-trained language models. While previous surveys have explored spatio-temporal data and methodologies separately, they have ignored a comprehensive examination of how foundation models are designed, selected, pre-trained, and adapted. As a result, the overall pipeline for spatio-temporal foundation models remains unclear. To bridge this gap, we innovatively provide an up-to-date review of previous spatio-temporal foundation models from the pipeline perspective. The pipeline begins with an introduction to different types of spatio-temporal data, followed by details of data preprocessing and embedding techniques. The pipeline then presents a novel data property taxonomy to divide existing methods according to data sources and dependencies, providing efficient and effective model design and selection for researchers. On this basis, we further illustrate the training objectives of primitive models, as well as the adaptation techniques of transferred models. Overall, our survey provides a clear and structured pipeline to understand the connection between core elements of spatio-temporal foundation models while guiding researchers to get started quickly. Additionally, we introduce emerging opportunities such as multi-objective training in the field of spatio-temporal foundation models, providing valuable insights for researchers and practitioners.