arXiv subjects

Feifei Li

Publications and source records attributed to Feifei Li.

At least 37 records · Page 2Linked to original sources

arXiv2026

SafeRoPE: Risk-specific Head-wise Embedding Rotation for Safe Generation in Rectified Flow Transformers

Recent Text-to-Image (T2I) models based on rectified-flow transformers (e.g., SD3, FLUX) achieve high generative fidelity but remain vulnerable to unsafe semantics, especially when triggered by multi-token interactions. Existing mitigation methods largely rely on fine-tuning or attention modulation for concept unlearning; however, their expensive computational overhead and design tailored to U-Net-based denoisers hinder direct adaptation to transformer-based diffusion models (e.g., MMDiT). In this paper, we conduct an in-depth analysis of the attention mechanism in MMDiT and find that unsafe semantics concentrate within interpretable, low-dimensional subspaces at head level, where a finite set of safety-critical heads is responsible for unsafe feature extraction. We further observe that perturbing the Rotary Positional Embedding (RoPE) applied to the query and key vectors can effectively modify some specific concepts in the generated images. Motivated by these insights, we propose SafeRoPE, a lightweight and fine-grained safe generation framework for MMDiT. Specifically, SafeRoPE first constructs head-wise unsafe subspaces by decomposing unsafe embeddings within safety-critical heads, and computes a Latent Risk Score (LRS) for each input vector via projection onto these subspaces. We then introduce head-wise RoPE perturbations that can suppress unsafe semantics without degrading benign content or image quality. SafeRoPE combines both head-wise LRS and RoPE perturbations to perform risk-specific head-wise rotation on query and key vector embeddings, enabling precise suppression of unsafe outputs while maintaining generation fidelity. Extensive experiments demonstrate that SafeRoPE achieves SOTA performance in balancing effective harmful content mitigation and utility preservation for safe generation of MMDiT. Codes are available at https://github.com/deng12yx/SafeRoPE.

cs.CV↗

arXiv2026

Why Database Manuals Are Not Enough: Efficient and Reliable Configuration Tuning for DBMSs via Code-Driven LLM Agents

Modern database management systems (DBMSs) expose hundreds of configuration knobs that critically influence performance. Existing automated tuning methods either adopt a data-driven paradigm, which incurs substantial overhead, or rely on manual-driven heuristics extracted from database documentation, which are often limited and overly generic. Motivated by the fact that the control logic of configuration knobs is inherently encoded in the DBMS source code, we argue that promising tuning strategies can be mined directly from the code, uncovering fine-grained insights grounded in system internals. To this end, we propose SysInsight, a code-driven database tuning system that automatically extracts fine-grained tuning knowledge from DBMS source code to accelerate and stabilize the tuning process. SysInsight combines static code analysis with LLM-based reasoning to identify knob-controlled execution paths and extract semantic tuning insights. These insights are then transformed into quantitative and verifiable tuning rules via association rule mining grounded in tuning observations. During online tuning, system diagnosis is applied to identify critical knobs, which are adjusted under the rule guidance. Evaluations demonstrate that compared to the SOTA baseline, SysInsight converges to the best configuration on average 7.11X faster while achieving a 19.9% performance improvement.

cs.DB↗

arXiv2026

CARROT: A Learned Cost-Constrained Retrieval Optimization System for RAG

Large Language Models (LLMs) have demonstrated impressive ability in generation and reasoning tasks but struggle with handling up-to-date knowledge, leading to inaccuracies or hallucinations. Retrieval-Augmented Generation (RAG) mitigates this by retrieving and incorporating external knowledge into input prompts. In particular, due to LLMs' context window limitations and long-context hallucinations, only the most relevant "chunks" are retrieved. However, current RAG systems face three key challenges: (1) chunks are often retrieved independently without considering their relationships, such as redundancy and ordering; (2) the utility of chunks is non-monotonic, as adding more chunks can degrade quality; and (3) retrieval strategies fail to adapt to the unique characteristics of different queries. To overcome these challenges, we design a cost-constrained retrieval optimization framework for RAG. We adopt a Monte Carlo Tree Search (MCTS) based strategy to find the optimal chunk combination order, which considers the chunks' correlations. In addition, to address the non-monotonicity of chunk utility, instead of treating budget exhaustion as the termination condition, we design a utility computation strategy to identify the optimal chunk combination without necessarily exhausting the budget. Furthermore, we propose a configuration agent that predicts optimal configurations for each query domain, improving our framework's adaptability and efficiency. Experimental results demonstrate up to a 30% improvement over baseline models, highlighting the framework's effectiveness, scalability, and suitability. Our source code has been released at https://github.com/wang0702/CARROT.

cs.DB↗

arXiv2026

Efficient Graph Embedding at Scale: Optimizing CPU-GPU-SSD Integration

Graph embeddings map graph nodes to continuous vectors and are foundational to community detection, recommendation, and many scientific applications. At billion-scale, however, existing graph embedding systems face a trade-off: they either rely on large in-memory footprints across many GPUs (limited scalability) or repeatedly stream data from disk (incurring severe I/O overhead and low GPU utilization). In this paper, we propose Legend, a lightweight heterogeneous system for graph embedding that systematically redesigns data management across CPU, GPU, and NVMe SSD resources. Legend combines three practical ideas: (1) a prefetch-friendly embedding-loading order that lets GPUs efficiently prefetch necessary embeddings directly from NVMe SSD with low I/O amplification; (2) a high-throughput GPU-SSD direct-access driver tuned for the access patterns of embedding training; and (3) a customized parallel execution strategy that maximizes GPU utilization. Together, these components let Legend store and stream vast embedding data without overprovisioning GPU memory or suffering I/O stalls. Extensive experiments on billion-scale graphs demonstrate that Legend speeds up end-to-end workloads by up to 4.8x versus state-of-the-art systems, and matches their performance on the largest workloads while using only one quarter of the GPUs.

cs.DC↗

arXiv2026

PoI: A Filter to Extract Pixel of Interest from Novel Views for Scene Coordinate Regression

Neural View Synthesis (NVS) techniques such as NeRF and 3D Gaussian Splatting (3DGS) have enabled photorealistic rendering from novel viewpoints and are increasingly used to augment training data for visual localization. However, these methods fundamentally rely on observed geometry and radiance; they interpolate existing information but cannot hallucinate unseen 3D structures or recover missing content under sparse or extreme viewpoints. As a result, rendered views often exhibit blur, structural distortion, or incomplete geometry. While such imperfections may be tolerated by Camera Pose Regression (CPR) methods, they severely degrade Scene Coordinate Regression (SCR), which requires accurate per-pixel 3D supervision. To address this limitation, we introduce PoI (Pixel-of-Interest), a framework that enables effective NVS augmentation for SCR-based localization. We first employ 3DGS to render novel views and leverage a single-step diffusion model to refine them, allowing the synthesis of structurally plausible details beyond purely geometry-driven interpolation. However, even diffusion-refined views may contain unreliable pixels. Therefore, we propose a progressive pixel-level filtering strategy based on reprojection error to selectively retain trustworthy synthetic pixels during training while suppressing harmful ones. Extensive experiments on 7Scenes and Cambridge Landmarks demonstrate that our method consistently improves localization accuracy over strong SCR baselines and achieves state-of-the-art performance with competitive training efficiency. Our results reveal that, for SCR, the benefit of novel view augmentation depends not only on generative realism but also on explicit control of pixel-level reliability.

cs.CV↗

arXiv2026

Adaptive Multi-Objective Tiered Storage Configuration for KV Cache in LLM Service

The memory-for-computation paradigm of KV caching is essential for accelerating large language model (LLM) inference service, but limited GPU high-bandwidth memory (HBM) capacity motivates offloading the KV cache to cheaper external storage tiers. While this expands capacity, it introduces the challenge of dynamically managing heterogeneous storage resources to balance cost, throughput, and latency under varying workloads. We formulate this as a multi-objective optimization problem: identifying the Pareto frontier across these metrics within the storage configuration space. Using a high-fidelity end-to-end simulator, we observe that the objective functions are non-analytic and exhibit complex variable coupling, making the Pareto frontier difficult to approximate analytically. To obtain the frontier, we introduce Kareto, a KV-cache Adaptive REsource managemenT Optimizer. Kareto leverages a diminishing-return-guided pruning method to efficiently navigate the large configuration space and approximate the Pareto frontier. Additionally, it incorporates a fine-grained adaptive tuner that uses eviction policies in tier storage and KV block access patterns for group-specific cache management, improving cache efficiency. Experiments on real-world traces show that Kareto adapts to workload and can identify configurations of better cost efficiency, covering static strategies. Compared to the fixed setup with 1024 GB DRAM, Kareto can improve throughput by up to 9.3%, or reduce latency by up to 58.3%, or lower cost by up to 20.2% under respective optimization objectives.

cs.AR↗

arXiv2026

Detecting Logic Bugs of Join Optimizations in DBMS

Generation-based testing techniques have shown their effectiveness in detecting logic bugs of DBMS, which are often caused by improper implementation of query optimizers. Nonetheless, existing generation-based debug tools are limited to single-table queries and there is a substantial research gap regarding multi-table queries with join operators. In this paper, we propose TQS, a novel testing framework targeted at detecting logic bugs derived by queries involving multi-table joins. Given a target DBMS, TQS achieves the goal with two key components: Data-guided Schema and Query Generation (DSG) and Knowledge-guided Query Space Exploration (KQE). DSG addresses the key challenge of multi-table query debugging: how to generate ground-truth (query, result) pairs for verification. It adopts the database normalization technique to generate a testing schema and maintains a bitmap index for result tracking. To improve debug efficiency, DSG also artificially inserts some noises into the generated data. To avoid repetitive query space search, KQE forms the problem as isomorphic graph set discovery and combines the graph embedding and weighted random walk for query generation. We evaluated TQS on four popular DBMSs: MySQL, MariaDB, TiDB and the gray release of an industry-leading cloud-native database, anonymized as X-DB. Experimental results show that TQS is effective in finding logic bugs of join optimization in database management systems. It successfully detected 115 bugs within 24 hours, including 31 bugs in MySQL, 30 in MariaDB, 31 in TiDB, and 23 in X-DB respectively.

cs.DB↗

arXiv2026

Context-Free Grammar Inference for Complex Programming Languages in Black Box Settings

Grammar inference for complex programming languages remains a significant challenge, as existing approaches fail to scale to real world datasets within practical time constraints. In our experiments, none of the state-of-the-art tools, including Arvada, Treevada and Kedavra were able to infer grammars for complex languages such as C, C++, and Java within 48 hours. Arvada and Treevada perform grammar inference directly on full-length input examples, which proves inefficient for large files commonly found in such languages. While Kedavra introduces data decomposition to create shorter examples for grammar inference, its lexical analysis still relies on the original inputs. Additionally, its strict no-overgeneralization constraint limits the construction of complex grammars. To overcome these limitations, we propose Crucio, which builds a decomposition forest to extract short examples for lexical and grammar inference via a distributional matrix. Experimental results show that Crucio is the only method capable of successfully inferring grammars for complex programming languages (where the number of nonterminals is up to 23x greater than in prior benchmarks) within reasonable time limits. On the prior simple benchmark, Crucio achieves an average recall improvement of 1.37x and 1.19x over Treevada and Kedavra, respectively, and improves F1 scores by 1.21x and 1.13x.

cs.PL↗

arXiv2026

HiMeS: Hippocampus-inspired Memory System for Personalized AI Assistants

Large language models (LLMs) power many interactive systems such as chatbots, customer-service agents, and personal assistants. In knowledge-intensive scenarios requiring user-specific personalization, conventional retrieval-augmented generation (RAG) pipelines exhibit limited memory capacity and insufficient coordination between retrieval mechanisms and user-specific conversational history, leading to redundant clarification, irrelevant documents, and degraded user experience. Inspired by the hippocampus-neocortex memory mechanism, we propose HiMeS, an AI-assistant architecture that fuses short-term and long-term memory. Our contributions are fourfold: (1) A short-term memory extractor is trained end-to-end with reinforcement learning to compress recent dialogue and proactively pre-retrieve documents from the knowledge base, emulating the cooperative interaction between the hippocampus and prefrontal cortex. (2) A partitioned long-term memory network stores user-specific information and re-ranks retrieved documents, simulating distributed cortical storage and memory reactivation. (3) On a real-world industrial dataset, HiMeS significantly outperforms a cascaded RAG baseline on question-answering quality. (4) Ablation studies confirm the necessity of both memory modules and suggest a practical path toward more reliable, context-aware, user-customized LLM-based assistants.

cs.AI↗

arXiv2025

SmartSight: Mitigating Hallucination in Video-LLMs Without Compromising Video Understanding via Temporal Attention Collapse

Despite Video Large Language Models having rapidly advanced in recent years, perceptual hallucinations pose a substantial safety risk, which severely restricts their real-world applicability. While several methods for hallucination mitigation have been proposed, they often compromise the model's capacity for video understanding and reasoning. In this work, we propose SmartSight, a pioneering step to address this issue in a training-free manner by leveraging the model's own introspective capabilities. Specifically, SmartSight generates multiple candidate responses to uncover low-hallucinated outputs that are often obscured by standard greedy decoding. It assesses the hallucination of each response using the Temporal Attention Collapse score, which measures whether the model over-focuses on trivial temporal regions of the input video when generating the response. To improve efficiency, SmartSight identifies the Visual Attention Vanishing point, enabling more accurate hallucination estimation and early termination of hallucinated responses, leading to a substantial reduction in decoding cost. Experiments show that SmartSight substantially lowers hallucinations for Qwen2.5-VL-7B by 10.59% on VRIPT-HAL, while simultaneously enhancing video understanding and reasoning, boosting performance on VideoMMMU by up to 8.86%. These results highlight SmartSight's effectiveness in improving the reliability of open-source Video-LLMs.

cs.CV↗

arXiv2025

Beluga: A CXL-Based Memory Architecture for Scalable and Efficient LLM KVCache Management

The rapid increase in LLM model sizes and the growing demand for long-context inference have made memory a critical bottleneck in GPU-accelerated serving systems. Although high-bandwidth memory (HBM) on GPUs offers fast access, its limited capacity necessitates reliance on host memory (CPU DRAM) to support larger working sets such as the KVCache. However, the maximum DRAM capacity is constrained by the limited number of memory channels per CPU socket. To overcome this limitation, current systems often adopt RDMA-based disaggregated memory pools, which introduce significant challenges including high access latency, complex communication protocols, and synchronization overhead. Fortunately, the emerging CXL technology introduces new opportunities in KVCache design. In this paper, we propose Beluga, a novel memory architecture that enables GPUs and CPUs to access a shared, large-scale memory pool through CXL switches. By supporting native load/store access semantics over the CXL fabric, our design delivers near-local memory latency, while reducing programming complexity and minimizing synchronization overhead. We conduct a systematic characterization of a commercial CXL switch-based memory pool and propose a set of design guidelines. Based on Beluga, we design and implement Beluga-KVCache, a system tailored for managing the large-scale KVCache in LLM inference. Beluga-KVCache achieves an 89.6% reduction in Time-To-First-Token (TTFT) and 7.35x throughput improvement in the vLLM inference engine compared to RDMA-based solutions. To the best of our knowledge, Beluga is the first system that enables GPUs to directly access large-scale memory pools through CXL switches, marking a significant step toward low-latency, shared access to vast memory resources by GPUs.

cs.DC↗

arXiv2025

PolarStore: High-Performance Data Compression for Large-Scale Cloud-Native Databases

In recent years, resource elasticity and cost optimization have become essential for RDBMSs. While cloud-native RDBMSs provide elastic computing resources via disaggregated computing and storage, storage costs remain a critical user concern. Consequently, data compression emerges as an effective strategy to reduce storage costs. However, existing compression approaches in RDBMSs present a stark trade-off: software-based approaches incur significant performance overheads, while hardware-based alternatives lack the flexibility required for diverse database workloads. In this paper, we present PolarStore, a compressed shared storage system for cloud-native RDBMSs. PolarStore employs a dual-layer compression mechanism that combines in-storage compression in PolarCSD hardware with lightweight compression in software. This design leverages the strengths of both approaches. PolarStore also incorporates database-oriented optimizations to maintain high performance on critical I/O paths. Drawing from large-scale deployment experiences, we also introduce hardware improvements for PolarCSD to ensure host-level stability and propose a compression-aware scheduling scheme to improve cluster-level space efficiency. PolarStore is currently deployed on thousands of storage servers within PolarDB, managing over 100 PB of data. It achieves a compression ratio of 3.55 and reduces storage costs by approximately 60%. Remarkably, these savings are achieved while maintaining performance comparable to uncompressed clusters.

cs.DC↗

arXiv2025

Improving Hierarchical Representations of Vectorized HD Maps with Perspective Clues

The construction of vectorized High-Definition (HD) maps from onboard surround-view cameras has become a significant focus in autonomous driving. However, current map vector estimation pipelines face two key limitations: input-agnostic queries struggle to capture complex map structures, and the view transformation leads to information loss. These issues often result in inaccurate shape restoration or missing instances in map predictions. To address this concern, we propose a novel approach, namely \textbf{PerCMap}, which explicitly exploits clues from perspective-view features at both instance and point level. Specifically, at instance level, we propose Cross-view Instance Activation (CIA) to activate instance queries across surround-view images, thereby helping the model recover the instance attributes of map vectors. At point level, we design Dual-view Point Embedding (DPE), which fuses features from both views to generate input-aware positional embeddings and improve the accuracy of point coordinate estimation. Extensive experiments on \textit{nuScenes} and \textit{Argoverse 2} demonstrate that PerCMap achieves strong and consistent performance across benchmarks, reaching 67.1 and 70.5 mAP, respectively.

cs.CV↗

arXiv2025

One Agent to Serve All: a Lite-Adaptive Stylized AI Assistant for Millions of Multi-Style Official Accounts

Conversational agents deployed in industrial-scale official account platforms must generate responses that are both contextually grounded and stylistically aligned-requirements that existing methods struggle to meet. Chain-of-thought (CoT) prompting induces significant latency due to multi-turn reasoning; per-account fine-tuning is computationally prohibitive; and long prompt-based methods degrade the model's ability to grasp injected context and style. In this paper, we propose WeStar, a lite-adaptive framework for stylized contextual question answering that scales to millions of official accounts. WeStar combines context-grounded generation via RAG with style-aware generation using Parametric RAG (PRAG), where LoRA modules are dynamically activated per style cluster. Our contributions are fourfold: (1) We introduce WeStar, a unified framework capable of serving large volumes of official accounts with minimal overhead. (2) We propose a multi-dimensional, cluster-based parameter sharing scheme that enables compact style representation while preserving stylistic diversity. (3) We develop a style-enhanced Direct Preference Optimization (SeDPO) method to optimize each style cluster's parameters for improved generation quality. (4) Experiments on a large-scale industrial dataset validate the effectiveness and efficiency of WeStar, underscoring its pracitical value in real-world deployment.

cs.CL↗

arXiv2025

InfoCons: Identifying Interpretable Critical Concepts in Point Clouds via Information Theory

Interpretability of point cloud (PC) models becomes imperative given their deployment in safety-critical scenarios such as autonomous vehicles. We focus on attributing PC model outputs to interpretable critical concepts, defined as meaningful subsets of the input point cloud. To enable human-understandable diagnostics of model failures, an ideal critical subset should be *faithful* (preserving points that causally influence predictions) and *conceptually coherent* (forming semantically meaningful structures that align with human perception). We propose InfoCons, an explanation framework that applies information-theoretic principles to decompose the point cloud into 3D concepts, enabling the examination of their causal effect on model predictions with learnable priors. We evaluate InfoCons on synthetic datasets for classification, comparing it qualitatively and quantitatively with four baselines. We further demonstrate its scalability and flexibility on two real-world datasets and in two applications that utilize critical scores of PC.

cs.LG↗

arXiv2025

Yannakakis+: Practical Acyclic Query Evaluation with Theoretical Guarantees

Acyclic conjunctive queries form the backbone of most analytical workloads, and have been extensively studied in the literature from both theoretical and practical angles. However, there is still a large divide between theory and practice. While the 40-year-old Yannakakis algorithm has strong theoretical running time guarantees, it has not been adopted in real systems due to its high hidden constant factor. In this paper, we strive to close this gap by proposing Yannakakis+, an improved version of the Yannakakis algorithm, which is more practically efficient while preserving its theoretical guarantees. Our experiments demonstrate that Yannakakis+ consistently outperforms the original Yannakakis algorithm by 2x to 5x across a wide range of queries and datasets. Another nice feature of our new algorithm is that it generates a traditional DAG query plan consisting of standard relational operators, allowing Yannakakis+ to be easily plugged into any standard SQL engine. Our system prototype currently supports four different SQL engines (DuckDB, PostgreSQL, SparkSQL, and AnalyticDB from Alibaba Cloud), and our experiments show that Yannakakis+ is able to deliver better performance than their native query plans on 160 out of the 162 queries tested, with an average speedup of 2.41x and a maximum speedup of 47,059x.

cs.DB↗

arXiv2025

Detect-and-Guide: Self-regulation of Diffusion Models for Safe Text-to-Image Generation via Guideline Token Optimization

Text-to-image diffusion models have achieved state-of-the-art results in synthesis tasks; however, there is a growing concern about their potential misuse in creating harmful content. To mitigate these risks, post-hoc model intervention techniques, such as concept unlearning and safety guidance, have been developed. However, fine-tuning model weights or adapting the hidden states of the diffusion model operates in an uninterpretable way, making it unclear which part of the intermediate variables is responsible for unsafe generation. These interventions severely affect the sampling trajectory when erasing harmful concepts from complex, multi-concept prompts, thus hindering their practical use in real-world settings. In this work, we propose the safe generation framework Detect-and-Guide (DAG), leveraging the internal knowledge of diffusion models to perform self-diagnosis and fine-grained self-regulation during the sampling process. DAG first detects harmful concepts from noisy latents using refined cross-attention maps of optimized tokens, then applies safety guidance with adaptive strength and editing regions to negate unsafe generation. The optimization only requires a small annotated dataset and can provide precise detection maps with generalizability and concept specificity. Moreover, DAG does not require fine-tuning of diffusion models, and therefore introduces no loss to their generation diversity. Experiments on erasing sexual content show that DAG achieves state-of-the-art safe generation performance, balancing harmfulness mitigation and text-following performance on multi-concept real-world prompts.

cs.CV↗

arXiv2025

Femur: A Flexible Framework for Fast and Secure Querying from Public Key-Value Store

With increasing demands for privacy, it becomes necessary to protect sensitive user query data when accessing public key-value databases. Existing Private Information Retrieval (PIR) schemes provide full security but suffer from poor scalability, limiting their applicability in large-scale deployment. We argue that in many real-world scenarios, a more practical solution should allow users to flexibly determine the privacy levels of their queries in a theoretically guided way, balancing security and performance based on specific needs. To formally provide provable guarantees, we introduce a novel concept of distance-based indistinguishability, which can facilitate users to comfortably relax their security requirements. We then design Femur, an efficient framework to securely query public key-value stores with flexible security and performance trade-offs. It uses a space-efficient learned index to convert query keys into storage locations, obfuscates these locations with extra noise provably derived by the distance-based indistinguishability theory, and sends the expanded range to the server. The server then adaptively utilizes the best scheme to retrieve data. We also propose a novel variable-range PIR scheme optimized for bandwidth-constrained environments. Experiments show that Femur outperforms the state-of-the-art designs even when ensuring the same full security level. When users are willing to relax their privacy requirements, Femur can further improve the performance gains to up to 163.9X, demonstrating an effective trade-off between security and performance.

cs.CR↗