Retrieval-augmented generation (RAG) improves large language model (LLM) answer quality by grounding generation in external evidence, but processing retrieved contexts makes the prefill stage a dominant serving cost. RAG cache fusion reduces this cost by reusing precomputed key-value (KV) caches for retrieved chunks and selectively recomputing tokens under the current prompt. Existing selectors, however, face a dilemma between quality and efficiency: fast query-agnostic or final-layer query-to-context selectors can miss request-relevant evidence, whereas full-view query-aware selectors require broad context and layer visibility before recomputation and therefore stall the layer-wise cache-fusion pipeline. We present QCFuse, a compressed-view query-aware selector for RAG cache fusion. QCFuse uses chunk-anchor query probing to condition user-query states on compact per-chunk anchors and critical-layer profiling to identify recomputation tokens without all-layer inspection. We implement QCFuse in SGLang and evaluate it on four open-weight LLMs across six datasets. QCFuse reaches full-prefill-level quality. At matched quality, QCFuse achieves an average prefill-time speedup of 1.7x over full prefill and 1.5x over ProphetKV, the strongest quality-preserving baseline.

The use of Large Language Models (LLMs) for querying relational data has given rise to relQuery, a workload pattern that applies templated LLM calls to structured tables. As relQuery services become more widely adopted in applications such as AI-powered spreadsheets, fast response times under concurrent query loads are increasingly important. Unfortunately, current LLM engines face severe latency bottlenecks from Head-of-Line (HoL) blocking across three comparable inference phases: waiting, core running, and tail running. Existing static priority scheduling methods only address HoL blocking during the waiting phase, leaving two critical problems unsolved. First, the absence of a priority update mechanism causes inaccurate prioritization and continued HoL blocking during core execution. Second, suboptimal prefill-decode batching exacerbates HoL blocking in tail execution and worsens latency trade-offs between running and waiting relQueries. To address these problems, we propose RelServe, an optimized LLM engine for low-latency relQuery serving. RelServe features two core innovations: a Dynamic Priority Updater that continuously adjusts priorities while minimizing overhead via statistical approximations, and an Adaptive Batch Arranger that quantitatively evaluates candidate prefill and decode batches to minimize projected average latency. Extensive experiments on four real-world datasets using LLMs ranging from 13B to 70B parameters show that RelServe reduces average serving latency by up to 3.1x compared to vLLM.

Dynamism is common in AI computation, e.g., the dynamic tensor shapes and the dynamic control flows in models. Due to the long compilation time, existing runtime compilation damages the model efficiency, while the offline compilers either suffer from the long compilation time and device memory footprint to cover all the possible execution instances of a dynamic model, or sacrifice optimization opportunities for usability. In this paper, we rethink the feasibility of runtime compilation for dynamic models and identify that the key for it to work is to speed up the compilation or hide the compilation overhead. To do this, we propose a real-time compiler, DVM. In DVM, we design a runtime operator compiler based on a bytecode virtual machine to perform effective and efficient compilation for each dynamic operator instance given its input. Specifically, instead of compiling programs into machine code, we encode the operator program into bytecode on the CPU and decode the bytecode into virtual instructions for direct execution on the NPU. Based on the runtime operator compiler, we further propose an operator fuser, which performs symbol-deduction-based fusion on static graphs and runtime fusion on dynamic graphs. Both pattern- and stacking-based fusion are supported to increase fusion opportunities. Evaluation on operators, subgraphs, and models shows that, compared with TorchInductor, PyTorch-eager and MindSpore-graph-O0, we are up to 11.77× better in terms of the operator/model efficiency and up to 5 orders of magnitude faster in terms of the maximum compilation time.

Knowledge-Based Visual Question Answering (KB-VQA) relies on retrieval from large-scale knowledge bases, yet this stage is often treated simplistically. Existing methods typically adopt single-view indexing or naive multi-view fusion, leading to systematic coverage gaps. In this work, we demonstrate that different views exhibit strong complementarity in retrieval. Motivated by this observation, we propose SCAR, a Structure-aware Cross-View Retrieval framework that exploits cross-view structural complementarity at inference time without additional training. SCAR enhances retrieval via structure-aware similarity propagation within each view and explicit cross-view redundancy regulation. Experiments on multiple KB-VQA benchmarks demonstrate that SCAR substantially improves retrieval recall, approaches retrieval coverage upper bounds, and consistently boosts end-to-end KB-VQA performance with negligible inference overhead.

Existing sparse attention and KV cache compression methods for long-context LLM inference typically apply fixed sparsity patterns or uniform budgets across all attention heads, overlooking the substantial variation in attention behavior among heads and contexts. We observe two distinct entropy patterns among attention heads: Rigid Heads, whose entropy stays near zero across input segments, and Dynamic Heads, whose entropy fluctuates significantly. Crucially, the distribution of these types is context-dependent and cannot be predetermined offline. We therefore propose EntropyInfer, a training-free framework that uses attention entropy to adaptively allocate compute at the granularity of individual heads and segments during prefilling. For decoding, we introduce a latent KV cache compression scheme that leverages generated output tokens, rather than prefill tokens alone, to identify and retain the most critical cache entries. Extensive experiments on Llama, Qwen and openPangu model series show that EntropyInfer consistently outperforms baselines including SnapKV, AdaKV, and CritiPrefill, achieving up to 2.39× end-to-end speedup beyond 100k tokens with minimal quality degradation compared to full attention. The code is released in https://github.com/SHA-4096/EntropyInfer.

Autonomous AI research has advanced rapidly, but long-horizon ML research engineering remains difficult: agents must sustain coherent progress across task comprehension, environment setup, implementation, experimentation, and debugging over hours or days. We introduce AiScientist, a system for autonomous long-horizon engineering for ML research built on a simple principle: strong long-horizon performance requires both structured orchestration and durable state continuity. To this end, AiScientist combines hierarchical orchestration with a permission-scoped File-as-Bus workspace: a top-level Orchestrator maintains stage-level control through concise summaries and a workspace map, while specialized agents repeatedly re-ground on durable artifacts such as analyses, plans, code, and experimental evidence rather than relying primarily on conversational handoffs, yielding thin control over thick state. Across two complementary benchmarks, AiScientist improves PaperBench score by 10.54 points on average over the best matched baseline and achieves 81.82 Any Medal

High-quality annotated datasets are crucial for advancing machine learning in medical image analysis. However, a critical gap exists: most datasets either offer a single, clean ground truth, which hides real-world expert disagreement, or they provide multiple annotations without a separate gold standard for objective evaluation. To bridge this gap, we introduce CytoCrowd, a new public benchmark for cytology analysis. The dataset features 446 high-resolution images, each with two key components: (1) raw, conflicting annotations from four independent pathologists, and (2) a separate, high-quality gold-standard ground truth established by a senior expert. This dual structure makes CytoCrowd a versatile resource. It serves as a benchmark for standard computer vision tasks, such as object detection and classification, using the ground truth. Simultaneously, it provides a realistic testbed for evaluating annotation aggregation algorithms that must resolve expert disagreements. We provide comprehensive baseline results for both tasks. Our experiments demonstrate the challenges presented by CytoCrowd and establish its value as a resource for developing the next generation of models for medical image analysis.

Large-scale crowdsourcing with high-quality results, such as online surveys and data labeling tasks, is in wide demand. Existing techniques on task assignment result quality optimization, however, have addressed only a part of optimization space that does not necessarily reflect the real-world problem. For example, in reality, the proper worker set for task assignment and the workers’ submissions to the task are non-obvious in advance, which is not considered by existing techniques. Hence, this paper discusses a task assignment algorithm that dynamically probes the proper worker set(s) and worker submission models for the task to optimize task assignments in terms of quality, time, and expense. Specifically, we introduce CrowdBwO (Crowd Bandit with Optimization), a novel multi-armed bandit algorithm that is based on batched bandits and bandits with Knapsack and incorporates worker submission models. CrowdBwO dynamically determines and utilizes proper worker set(s) and worker submission models for each task under uncertainty to achieve high-performance crowdsourcing. We conducted extensive experiments with synthetic workers and real workers to evaluate CrowdBwO in two specific problem settings. Our extensive experimental results demonstrate that CrowdBwO is significant for real-world crowdsourcing and has a high performance.

Large Language Models (LLMs) excel at various reasoning tasks but still encounter challenges such as hallucination and factual inconsistency in knowledge-intensive tasks, primarily due to a lack of external knowledge and factual verification. These challenges could be mitigated by leveraging knowledge graphs (KGs) to support more reliable LLM reasoning. However, existing KG-augmented LLM frameworks still rely on static integration mechanisms that cannot adjust reasoning in response to evolving context and retrieved evidence, resulting in error propagation and incomplete reasoning. To alleviate these issues, we propose **V**erify-**o**n-**G**raph (**VoG**), a scalable and model-agnostic framework to enhance LLM reasoning via iterative retrieval, stepwise verification, and adaptive revision. Besides performing KG retrieval guided by an initially generated reasoning plan, VoG iteratively verifies and revises the reasoning plan, correcting intermediate errors in consideration of the varying contextual conditions. During plan revision, VoG leverages a context-aware multi-armed bandit strategy, guided by reward signals that capture uncertainty and semantic consistency, to enhance the alignment between the reasoning plan and retrieved evidence in a more adaptive and reliable way. Experimental results across three benchmark datasets show that VoG consistently improves both reasoning accuracy and efficiency. Our code is available at https://anonymous.4open.science/r/VoG-132C/.

Graph Neural Networks (GNNs) have become essential for solving graph-level tasks, such as classification and regression, across diverse domains including social networks and biology. However, existing GNNs struggle with the expressivity that captures complex structural patterns, and the generalization that ensures robust performance on diverse and noisy datasets. To address these challenges, we propose a novel GNN model that integrates a k-path rooted subgraph encoder, an adaptive graph contrastive learning approach, and a consistency-aware loss. The k-path rooted subgraph encoder enhances expressivity by capturing and distinguishing intricate substructures, with theoretical guarantees for counting paths and cycles. The adaptive graph contrastive learning framework improves generalization by generating domain-aware graph augmentations based on edge importance, while the consistency-aware loss ensures task-relevant properties are preserved across augmented views. Extensive experiments on 26 datasets spanning graph classification, regression, and realistic scenarios such as noise, class imbalance, and few-shot learning show that our model achieves superior performance against 18 state-of-the-art GNN models in both effectiveness and efficiency. The code is released in https://anonymous.4open.science/r/GEGNN .

The digitization of pathology slide has led to the proliferation of Whole Slide Image (WSI), creating a growing demand for efficient image retrieval systems that can enhance diagnostic accuracy and medical research. However, existing methods are often constrained by coarse slide-level summaries or patch-based searches, which fail to capture the hierarchical context of WSI structures and are computationally inefficient for region queries. To overcome these limitations, we propose HIRES, a Hierarchical Retrieval framework for flexible Whole Slide Image region search. HIRES features a novel hierarchical vector indexing strategy that models the pyramidal structure of WSIs, preserving both spatial relationships across magnifications and semantic similarities within each level. To address the computational demands of this complex search space, we pair this index with a GPU-accelerated retrieval pipeline. Experimental results show that our method achieves a 590 & times; speedup in query processing, and also significantly improves the retrieval quality on large WSI datasets compared to existing works.

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding and generation. However, their immense number of parameters and complex transformer-based architectures result in significant resource demands and computational complexity during training, making it challenging to optimize them efficiently on large datasets. To reduce training costs while preserving performance, researchers have investigated coreset selection techniques, which aim to identify small, representative subsets of the entire training dataset to accelerate LLM training. However, existing coreset selection methods fail to adapt to the dynamic nature of LLM training and often struggle with scalability for models of this size. To address these limitations, we propose a graph-guided adaptive and dynamic coreset selection framework for LLMs, namely GRACE. GRACE dynamically constructs and updates coresets by combining representation diversity with gradient-based importance metrics, ensuring both informativeness and efficiency. To mitigate the computational cost of frequent updates, GRACE leverages a k-NN graph-based propagation mechanism and selectively updates scores and embeddings, adapting to evolving training dynamics. Extensive experiments on three benchmarks demonstrate that GRACE significantly improves training efficiency and downstream performance across diverse LLMs and tasks.

The ongoing digitization has led to a proliferation of time-series data streams that monitor a variety of processes, from which valuable insights may be obtained. Further, the emergence of successful foundational language models begs the question of whether it is possible to achieve time-series models with the foundational properties of handling multiple tasks, while being sufficiently lightweight to allow real-time data stream processing. Existing foundational time-series models are often large and only effective in offline settings without stringent time and computational constraints, and where repeated model calibration is not needed. However, when applied to data streams, these models are ineffective due to their size and lack of support for continual calibration, which compromise their ability to deliver accurate real-time responses, their durability, and their deployability in hardware-limited settings. We propose TimeBlocks to enable versatile time-series processing by facilitating the efficient building of lightweight models suitable for multiple tasks under variable conditions. In particular, the method maintains a pool of interchangeable and modular model blocks that can be used to construct new time-series models. When presented with specific time-series data, a routing strategy iteratively selects the most suitable blocks to construct a lightweight and accurate model for the data. We equip TimeBlocks with a method called StreamCore to build a representative small subset of the data stream, which preserves a guaranteed approximation of the stream over time, enabling continual model calibration. An experimental study on multiple data sets and covering multiple tasks shows that TimeBlocks enables to build models capable of outperforming existing baselines.

Pedestrian trajectory prediction plays a vital role in autonomous driving and intelligent surveillance systems. Graph neural networks (GNNs) have shown remarkable effectiveness in this task by explicitly modeling social interactions among pedestrians. However, existing methods suffer from two key limitations. First, they face difficulty in balancing the reduction in redundant connections with the preservation of critical interaction relationships in spatial graph construction. Second, higher-order graph convolution methods lack adaptability to varying crowd densities. To address these limitations, we propose a pedestrian trajectory prediction method based on Delaunay triangulation and density-adaptive higher-order graph convolution. First, we leverage Delaunay triangulation to construct a sparse, geometrically principled adjacency structure for spatial interaction graphs, which effectively eliminates redundant connections while preserving essential proximity relationships. Second, we design a density-adaptive order selection mechanism that dynamically adjusts the graph convolution order according to pedestrian density. Experiments on the ETH/UCY datasets show that our method achieves 5.6% and 9.4% reductions in average displacement error (ADE) and final displacement error (FDE), respectively, compared with the recent graph convolution-based method DSTIGCN, demonstrating the effectiveness of the proposed approach.

Graph Neural Networks (GNNs) have achieved great success in many applications, and mini-batch training has become the de facto standard for training GNNs on large-scale graphs. When developing mini-batch GNN training systems on CPU-GPU platforms, existing dedicated batching systems adopt a static workload-processor binding strategy, where the costly mini-batch preparation workload is exclusively assigned to either the CPU or the GPU. This leads to low utilization of CPU cores, PCIe bandwidth, and GPU computing capability, resulting in suboptimal training efficiency. To address this problem, we develop MorphGL, a novel GNN training system featuring a collective batching design. MorphGL adaptively dispatches the mini-batch preparation workload to both the CPU and GPU, ensuring that the workload distribution aligns with the CPU-GPU setup of the running machine for optimal efficiency. To maximize resource utilization, MorphGL employs the Dual-Buffer Scheduling algorithm to collectively schedule training stages across the CPU, PCIe, and GPU. Extensive experiments on three large real-world graphs with billions of edges and four machines with representative CPU-GPU configurations demonstrate that MorphGL consistently outperforms state-of-the-art GNN training systems, achieving up to 2.76x and 2.2x speedup over SALIENT and DUCATI, respectively.

Logs, generated from modern computational systems such as cloud servers or DBMS, are the primary indicator of system states and thus have drawn significant attention from researchers. One of its key tasks is log anomaly detection, aiming to discover anomalous signals that can subsequently imply errors in systems. Conventionally, the major challenge of such a detection task is the insufficiency of well-labeled logs, which require unaffordable human resources. To tackle this issue, recent works have adopted the large language models (LLMs) as zeroshot label-free log anomaly detectors. However, these solutions necessitate direct deployment of LLM instances for downstream tasks, incurring substantial computational costs. Such costs limit their applications in resource-constrained scenarios, which motivates us to reposition the LLMs from direct detectors to training assistants and propose MaidLog. Specifically, MaidLog comprises an LLM-assisted pseudo-labels assignment workflow that automatically generates high-quality labels at training, enabling it to run entirely without manual labels. Based on them, we propose a lightweight detector designed for generalizability, where a well-trained detector is even applicable to downstream systems without any post-training. Once the training is finished, the downstream detection is efficient and only detector-based, without any involvement of LLM. Extensive experiments show that MaidLog effectively handles challenging real-world scenarios. With only a few unlabeled entries, MaidLog achieves comparable performance with a $25,000 \mathrm{x}$ speedup over SOTA LLM-centric solutions in out-of-domain systems, demonstrating its applicability on resource-constrained systems.

Large Language Models (LLMs) have shown strong capabilities in open-domain question answering (QA), but deploying them in real-world online systems introduces critical challenges. These include: (1) handling both simple and complex queries with appropriate levels of reasoning, (2) minimizing latency without compromising answer quality, and (3) maintaining answer consistency under evolving and noisy retrieval contexts. To address these challenges, we propose Retain-to-Refine ( (RR)-R-2), an adaptive agent-based QA framework designed for practical deployment. (RR)-R-2 integrates a Query Critic Agent (QCA) to assess query difficulty and route it accordingly: simple queries are answered directly using fast, prompt-based LLM calls, while complex queries are handled by a Memory Augmented Agent (MAA). MAA performs iterative reasoning guided by a unique long-short memory mechanism. Long-term memory retains and consolidates stable, core facts to ground the reasoning process, while short-term memory identifies transient information gaps to formulate highly focused subsequent queries. To ensure evidence quality, a Supervised Retrospection module validates and filters retrieved documents at each step. This agent-based design enables R2R to dynamically allocate computation based on question complexity, reducing unnecessary overhead while preserving high-quality answers when multi-step reasoning or external knowledge is required. Extensive evaluations across various settings and datasets demonstrate that the efficiency of (RR)-R-2 across diverse question types. In online settings, (RR)-R-2 delivers substantial gains in both response quality and efficiency, making it well-suited for large-scale industrial deployment in real-time QA services.