arXiv subjects

Zihe Song

Publications and source records attributed to Zihe Song.

At least 19 recordsLinked to original sources

arXiv2026

An Investigation on Numerical Bugs in GPU Programs Towards Automated Bug Detection

General-purpose graphics processing unit (GPU) computing has emerged as a leading parallel computing paradigm, offering significant performance gains in various domains such as scientific computing and deep learning. However, GPU programs are susceptible to numerical bugs, which can lead to incorrect results or crashes. These bugs are difficult to detect, debug, and fix due to their dependence on specific input values or types and the absence of reliable error-checking mechanisms and oracles. Additionally, the unique programming conventions of GPUs complicate identifying the root causes of bugs, while fixing them requires domain-specific knowledge of GPU computing and numerical libraries. Therefore, understanding the characteristics of GPU numerical bugs (GPU-NBs) is crucial for developing effective solutions. In this paper, we conduct a comprehensive study of GPU-NBs by analyzing 397 real-world bug samples from GitHub. We identify common root causes, symptoms, input patterns, test oracles that trigger these bugs and the strategies used to fix them. We also present GPU-NBDetect, a preliminary tool designed to detect numerical bugs across six distinct bug categories. GPU-NBDetect detected a total of 226 bugs across 186 mathematical functions in four libraries, with 60 of the bugs confirmed by developers. Our findings lay the groundwork for developing detection and prevention techniques for GPU-NBs and offer insights for building more effective debugging and auto-repair tool.

cs.SE↗

arXiv2026

SPECMEM: Accelerating Agentic Tool Calling via Live Memory Management

Tool-calling agents are mostly optimized for accuracy; in deployment, however, their growing cost is latency: every tool call is decoded autoregressively. Speculative decoding accelerates this without changing the output, and retrieved past calls make strong drafts under real use, where each user's behavior tends to repeat. Yet existing systems keep no cross-session memory, or build the draft store once and never update it. We propose SpecMem, a tool-call memory kept alive and managed throughout deployment: verified calls are written back online into bounded per-user stores with LRU eviction, and a side-effect gate defers irreversible calls until verification confirms. On multi-session, multi-user query streams served by three open-weight models, a frozen store steadily loses acceptance while SpecMem holds it: a 27% gap in mean accepted tokens (+11% to +34% across benchmarks, tracking how much novel work arrives after the store is built) and up to 1.74x wall-clock speedup. Ablations attribute the gain almost entirely to keeping the store live; local per-user scoping adds no acceptance but closes a measured cross-user leakage channel, and gating on declared side effects, rather than on retrieval confidence, is what eliminates irreversible mistakes.

cs.CL↗

arXiv2026

TileCas: Cascading Kernel Development Across Programming Abstractions

Programming abstractions have traditionally been designed for humans, but code generation is increasingly performed by models and autonomous agents. We investigate abstractions for machine-driven programming through GPU and AI-accelerator kernel generation. High-level domain-specific languages (DSLs) simplify search by delegating implementation details to compilers, while lower-level representations provide finer-grained control at the cost of a larger and more complex search space. In this paper, we present TileCas, a unified, pluggable interface that enables coding agents to develop kernels across multiple editable abstraction levels. Rather than committing to a single representation, TileCas allows an agent to transfer and refine candidates across levels through Cross-Level Diagnosis, which maps correctness and performance feedback to actionable controls at different abstraction levels; Abstraction Scheduling, which selects both the candidate and representation from which subsequent refinement should proceed; and a Heterogeneous Search Tree, which preserves and revisits alternative search trajectories spanning multiple representations. We integrate TileCas with two code agents AKO and KDA, and evaluate both integrations on 50 end-to-end KernelBench Level 3 workloads across two accelerator platforms and two language models. Giving equal weight to the four accelerator-model configurations, AKO with TileCas achieves an arithmetic-mean speedup of $4.84\times$, compared with $3.75\times$ for the strongest evaluated baseline without TileCas, an improvement of 29.0%. Integrating the same interface with KDA improves its mean speedup by 69.2%, with gains in every configuration. These results show that allowing coding agents to navigate across programming abstractions can improve machine-driven software development and can be reused across different base workflows.

cs.SE↗

arXiv2026

Characterizing Real-World Bugs in Tile Programs for Automated Bug Detection

Tile-based programming frameworks are increasingly adopted to write high-performance GPU kernels in domains such as deep learning and scientific computing. While these frameworks enhance productivity and hardware utilization, their multi-stage compilation pipelines introduce distinct code generation bugs that are tightly coupled to input shapes, data types, and backend targets. These bugs often manifest as silent wrong results or performance issues, making them difficult to detect using existing compiler testing tools. Additionally, the unique programming conventions of tile domain-specific languages complicate root cause identification, while fixing such bugs demands specialized knowledge of tile abstractions and compilation pipelines. Despite the growing adoption of tile-based systems, their code generation bugs remain largely unexplored. This paper presents the first systematic study of tile-program code generation bugs. We curate 401 bug reports from GitHub and identify 301 tile-program codegen bugs for analysis, characterizing their root causes and symptoms, the input patterns that trigger them, the test oracles that detect them, and the strategies for fixing these bugs. Our study provides foundational insights for building debugging, testing, and repair tools tailored to tile-based compiler infrastructures.

cs.SE↗

arXiv2026

BlockServe: Block-Grained Continuous Batching for High-Throughput Diffusion LLM Serving

Efficient serving of diffusion large language models (dLLMs) is hindered by convergence heterogeneity: when batching multiple requests, different sequences converge at different rates, causing faster requests to stall behind slower stragglers and introducing compute bubbles and tail latency. We present BlockServe, a continuous batching framework that integrates block-grained scheduling -- immediately evicting completed requests at block boundaries -- with mixed-state execution that extends dual cache and parallel decoding to heterogeneous batches via gather-scatter indexing. Furthermore, a compute-aware admission controller expands effective batch capacity through token-budgeted refill. On Dream and LLaDA across five benchmarks, BlockServe achieves 1.9--10.6$\times$ throughput over Fast-dLLM with comparable generation quality, establishing block-grained scheduling as a foundation for high-throughput offline dLLM inference.

cs.LG↗

arXiv2026

Why LLMs Hallucinate on Structured Knowledge: A Mechanistic Analysis of Reasoning over Linearized Representations

In many reasoning tasks, large language models (LLMs) rely on structured external knowledge, such as graphs and tables, which is typically linearized into sequential token representations. However, even when sufficient knowledge is available, LLMs can still produce hallucinated outputs, and the underlying mechanisms behind such failures remain poorly understood. We investigate these mechanisms and find that hallucinations arise from systematic internal dynamics rather than random noise. First, attention disproportionately concentrates toward shortcut-like structural cues rather than distributing across the full context. Second, feed-forward representations fail to ground the provided knowledge, causing the model to revert to parametric memory. Moreover, our results indicate that hallucination is consistently associated with failures in semantic grounding within feed-forward layers, while attention allocation exhibits greater task-dependent variability. Finally, we show that these mechanistic patterns generalize beyond single-hop graphs to multi-hop and tabular settings, enabling effective hallucination detection across structured knowledge formats.

cs.CL↗

arXiv2026

LLM-MemCluster: Empowering Large Language Models with Dynamic Memory for Text Clustering

Large Language Models (LLMs) are reshaping unsupervised learning by offering an unprecedented ability to perform text clustering based on their deep semantic understanding. However, their direct application is fundamentally limited by a lack of stateful memory for iterative refinement and the difficulty of managing cluster granularity. As a result, existing methods often rely on complex pipelines with external modules, sacrificing a truly end-to-end approach. We introduce LLM-MemCluster, a novel framework that reconceptualizes clustering as a fully LLM-native task. It leverages a Dynamic Memory to instill state awareness and a Dual-Prompt Strategy to enable the model to reason about and determine the number of clusters. Evaluated on several benchmark datasets, our tuning-free framework significantly and consistently outperforms strong baselines. LLM-MemCluster presents an effective, interpretable, and truly end-to-end paradigm for LLM-based text clustering.

cs.CL↗

arXiv2025

Detecting Hallucinations in Graph Retrieval-Augmented Generation via Attention Patterns and Semantic Alignment

Graph-based Retrieval-Augmented Generation (GraphRAG) enhances Large Language Models (LLMs) by incorporating external knowledge from linearized subgraphs retrieved from knowledge graphs. However, LLMs struggle to interpret the relational and topological information in these inputs, resulting in hallucinations that are inconsistent with the retrieved knowledge. To analyze how LLMs attend to and retain structured knowledge during generation, we propose two lightweight interpretability metrics: Path Reliance Degree (PRD), which measures over-reliance on shortest-path triples, and Semantic Alignment Score (SAS), which assesses how well the model's internal representations align with the retrieved knowledge. Through empirical analysis on a knowledge-based QA task, we identify failure patterns associated with over-reliance on salient paths and weak semantic grounding, as indicated by high PRD and low SAS scores. We further develop a lightweight post-hoc hallucination detector, Graph Grounding and Alignment (GGA), which outperforms strong semantic and confidence-based baselines across AUC and F1. By grounding hallucination analysis in mechanistic interpretability, our work offers insights into how structural limitations in LLMs contribute to hallucinations, informing the design of more reliable GraphRAG systems in the future.

cs.CL↗

arXiv2025

Topology-Aware Conformal Prediction for Stream Networks

Stream networks, a unique class of spatiotemporal graphs, exhibit complex directional flow constraints and evolving dependencies, making uncertainty quantification a critical yet challenging task. Traditional conformal prediction methods struggle in this setting due to the need for joint predictions across multiple interdependent locations and the intricate spatio-temporal dependencies inherent in stream networks. Existing approaches either neglect dependencies, leading to overly conservative predictions, or rely solely on data-driven estimations, failing to capture the rich topological structure of the network. To address these challenges, we propose Spatio-Temporal Adaptive Conformal Inference (\texttt{STACI}), a novel framework that integrates network topology and temporal dynamics into the conformal prediction framework. \texttt{STACI} introduces a topology-aware nonconformity score that respects directional flow constraints and dynamically adjusts prediction sets to account for temporal distributional shifts. We provide theoretical guarantees on the validity of our approach and demonstrate its superior performance on both synthetic and real-world datasets. Our results show that \texttt{STACI} effectively balances prediction efficiency and coverage, outperforming existing conformal prediction methods for stream networks.

stat.ML↗

arXiv2025

AppForge: From Assistant to Independent Developer -- Are GPTs Ready for Software Development?

Large language models (LLMs) have demonstrated remarkable capability in function-level code generation tasks. Unlike isolated functions, real-world applications demand reasoning over the entire software system: developers must orchestrate how different components interact, maintain consistency across states over time, and ensure the application behaves correctly within the lifecycle and framework constraints. Yet, no existing benchmark adequately evaluates whether LLMs can bridge this gap and construct entire software systems from scratch. To address this gap, we propose APPFORGE, a benchmark consisting of 101 software development problems drawn from real-world Android apps. Given a natural language specification detailing the app functionality, a language model is tasked with implementing the functionality into an Android app from scratch. Developing an Android app from scratch requires understanding and coordinating app states, lifecycle management, and asynchronous operations, calling for LLMs to generate context-aware, robust, and maintainable code. To construct APPFORGE, we design a multi-agent system to automatically summarize the main functionalities from app documents and navigate the app to synthesize test cases validating the functional correctness of app implementation. Following rigorous manual verification by Android development experts, APPFORGE incorporates the test cases within an automated evaluation framework that enables reproducible assessment without human intervention, making it easily adoptable for future research. Our evaluation on 12 flagship LLMs show that all evaluated models achieve low effectiveness, with the best-performing model (GPT-5) developing only 18.8% functionally correct applications, highlighting fundamental limitations in current models' ability to handle complex, multi-component software engineering challenges.

cs.SE↗

arXiv2025

SGCL: Unifying Self-Supervised and Supervised Learning for Graph Recommendation

Recommender systems (RecSys) are essential for online platforms, providing personalized suggestions to users within a vast sea of information. Self-supervised graph learning seeks to harness high-order collaborative filtering signals through unsupervised augmentation on the user-item bipartite graph, primarily leveraging a multi-task learning framework that includes both supervised recommendation loss and self-supervised contrastive loss. However, this separate design introduces additional graph convolution processes and creates inconsistencies in gradient directions due to disparate losses, resulting in prolonged training times and sub-optimal performance. In this study, we introduce a unified framework of Supervised Graph Contrastive Learning for recommendation (SGCL) to address these issues. SGCL uniquely combines the training of recommendation and unsupervised contrastive losses into a cohesive supervised contrastive learning loss, aligning both tasks within a single optimization direction for exceptionally fast training. Extensive experiments on three real-world datasets show that SGCL outperforms state-of-the-art methods, achieving superior accuracy and efficiency.

cs.IR↗

arXiv2025

Efficiency Robustness of Dynamic Deep Learning Systems

Deep Learning Systems (DLSs) are increasingly deployed in real-time applications, including those in resourceconstrained environments such as mobile and IoT devices. To address efficiency challenges, Dynamic Deep Learning Systems (DDLSs) adapt inference computation based on input complexity, reducing overhead. While this dynamic behavior improves efficiency, such behavior introduces new attack surfaces. In particular, efficiency adversarial attacks exploit these dynamic mechanisms to degrade system performance. This paper systematically explores efficiency robustness of DDLSs, presenting the first comprehensive taxonomy of efficiency attacks. We categorize these attacks based on three dynamic behaviors: (i) attacks on dynamic computations per inference, (ii) attacks on dynamic inference iterations, and (iii) attacks on dynamic output production for downstream tasks. Through an in-depth evaluation, we analyze adversarial strategies that target DDLSs efficiency and identify key challenges in securing these systems. In addition, we investigate existing defense mechanisms, demonstrating their limitations against increasingly popular efficiency attacks and the necessity for novel mitigation strategies to secure future adaptive DDLSs.

cs.LG↗

arXiv2025

Item Cluster-aware Prompt Learning for Session-based Recommendation

Session-based recommendation (SBR) aims to capture dynamic user preferences by analyzing item sequences within individual sessions. However, most existing approaches focus mainly on intra-session item relationships, neglecting the connections between items across different sessions (inter-session relationships), which limits their ability to fully capture complex item interactions. While some methods incorporate inter-session information, they often suffer from high computational costs, leading to longer training times and reduced efficiency. To address these challenges, we propose the CLIP-SBR (Cluster-aware Item Prompt learning for Session-Based Recommendation) framework. CLIP-SBR is composed of two modules: 1) an item relationship mining module that builds a global graph to effectively model both intra- and inter-session relationships, and 2) an item cluster-aware prompt learning module that uses soft prompts to integrate these relationships into SBR models efficiently. We evaluate CLIP-SBR across eight SBR models and three benchmark datasets, consistently demonstrating improved recommendation performance and establishing CLIP-SBR as a robust solution for session-based recommendation tasks.

cs.IR↗

arXiv2025

Can You Mimic Me? Exploring the Use of Android Record & Replay Tools in Debugging

Android User Interface (UI) testing is a critical research area due to the ubiquity of apps and the challenges faced by developers. Record and replay (R&R) tools facilitate manual and automated UI testing by recording UI actions to execute test scenarios and replay bugs. These tools typically support (i) regression testing, (ii) non-crashing functional bug reproduction, and (iii) crashing bug reproduction. However, prior work only examines these tools in fragmented settings, lacking a comprehensive evaluation across common use cases. We address this gap by conducting an empirical study on using R&R tools to record and replay non-crashing failures, crashing bugs, and feature-based user scenarios, and explore combining R&R with automated input generation (AIG) tools to replay crashing bugs. Our study involves one industrial and three academic R&R tools, 34 scenarios from 17 apps, 90 non-crashing failures from 42 apps, and 31 crashing bugs from 17 apps. Results show that 17% of scenarios, 38% of non-crashing bugs, and 44% of crashing bugs cannot be reliably recorded and replayed, mainly due to action interval resolution, API incompatibility, and Android tooling limitations. Our findings highlight key future research directions to enhance the practical application of R&R tools.

cs.SE↗

arXiv2025

An Empirical Analysis of Compatibility Issues for Industrial Mobile Games

Detecting and fixing compatibility issues is critical for mobile game development. The rapid evolution of mobile operating systems and device fragmentation make it challenging for developers to timely address these issues across diverse models. Undetected compatibility problems can severely impact user experience and cause financial loss to companies and players. However, mobile game testing remains highly challenging, and compatibility issues are largely underexplored by the research community. To bridge this gap, we conduct an empirical study on common compatibility issues in commercial mobile games. We select four active and representative games with well-documented bug reports, totaling over seven million lines of code and over 20,000 commits. We build a comprehensive dataset linking bugs and fixes, enabling investigation into prevalent symptoms, root causes, and fixing strategies. Through extensive manual analysis, we categorize the most common symptoms and root causes, and summarize the typical fixes for each category. Our findings provide practical guidance for developers and offer insights to inspire future research on testing and fixing compatibility issues in mobile games.

cs.SE↗

arXiv2025

Graph Neural Controlled Differential Equations For Collaborative Filtering

Graph Convolution Networks (GCNs) are widely considered state-of-the-art for recommendation systems. Several studies in the field of recommendation systems have attempted to apply collaborative filtering (CF) into the Neural ODE framework. These studies follow the same idea as LightGCN, which removes the weight matrix or with a discrete weight matrix. However, we argue that weight control is critical for neural ODE-based methods. The importance of weight in creating tailored graph convolution for each node is crucial, and employing a fixed/discrete weight means it cannot adjust over time within the ODE function. This rigidity in the graph convolution reduces its adaptability, consequently hindering the performance of recommendations. In this study, to create an optimal control for Neural ODE-based recommendation, we introduce a new method called Graph Neural Controlled Differential Equations for Collaborative Filtering (CDE-CF). Our method improves the performance of the Graph ODE-based method by incorporating weight control in a continuous manner. To evaluate our approach, we conducted experiments on various datasets. The results show that our method surpasses competing baselines, including GCNs-based models and state-of-the-art Graph ODE-based methods.

cs.IR↗

arXiv2024

Do We Really Need Graph Convolution During Training? Light Post-Training Graph-ODE for Efficient Recommendation

The efficiency and scalability of graph convolution networks (GCNs) in training recommender systems (RecSys) have been persistent concerns, hindering their deployment in real-world applications. This paper presents a critical examination of the necessity of graph convolutions during the training phase and introduces an innovative alternative: the Light Post-Training Graph Ordinary-Differential-Equation (LightGODE). Our investigation reveals that the benefits of GCNs are more pronounced during testing rather than training. Motivated by this, LightGODE utilizes a novel post-training graph convolution method that bypasses the computation-intensive message passing of GCNs and employs a non-parametric continuous graph ordinary-differential-equation (ODE) to dynamically model node representations. This approach drastically reduces training time while achieving fine-grained post-training graph convolution to avoid the distortion of the original training embedding space, termed the embedding discrepancy issue. We validate our model across several real-world datasets of different scales, demonstrating that LightGODE not only outperforms GCN-based models in terms of efficiency and effectiveness but also significantly mitigates the embedding discrepancy commonly associated with deeper graph convolution layers. Our LightGODE challenges the prevailing paradigms in RecSys training and suggests re-evaluating the role of graph convolutions, potentially guiding future developments of efficient large-scale graph-based RecSys.

cs.LG↗

arXiv2024

ImplicitAVE: An Open-Source Dataset and Multimodal LLMs Benchmark for Implicit Attribute Value Extraction

Existing datasets for attribute value extraction (AVE) predominantly focus on explicit attribute values while neglecting the implicit ones, lack product images, are often not publicly available, and lack an in-depth human inspection across diverse domains. To address these limitations, we present ImplicitAVE, the first, publicly available multimodal dataset for implicit attribute value extraction. ImplicitAVE, sourced from the MAVE dataset, is carefully curated and expanded to include implicit AVE and multimodality, resulting in a refined dataset of 68k training and 1.6k testing data across five domains. We also explore the application of multimodal large language models (MLLMs) to implicit AVE, establishing a comprehensive benchmark for MLLMs on the ImplicitAVE dataset. Six recent MLLMs with eleven variants are evaluated across diverse settings, revealing that implicit value extraction remains a challenging task for MLLMs. The contributions of this work include the development and release of ImplicitAVE, and the exploration and benchmarking of various MLLMs for implicit AVE, providing valuable insights and potential future research directions. Dataset and code are available at https://github.com/HenryPengZou/ImplicitAVE

cs.CV↗