arXiv subjects

Xingyu Fan

Publications and source records attributed to Xingyu Fan.

11 recordsLinked to original sources

arXiv2026

When History Lies: Evaluating and Improving Tool Use under Misleading Multi-Turn Histories

Tool-calling agents infer task state from accumulated dialogue and tool traces. In persistent interactions, however, historical traces may remain structurally valid and semantically plausible after they cease to be authoritative for the current request. We show that such history can hijack a policy the model already possesses: on Qwen3-1.7B, pollution flips 32.1% of decisions that are correct under the original trajectory and frequently induces reuse of corrupted entities or interface conventions. We introduce bench, a paired benchmark with synchronized Original, Polluted, and Oracle State views that preserve the system policy, current tools, latest request, and gold next action. Eleven gold-preserving interventions isolate failures in decision state, entity binding, and interface execution across complete calls and non-call decisions. We further propose ours, which transfers an Oracle-conditioned teacher policy to a student observing only polluted history through soft supervision on student-generated prefixes. On Qwen3-1.7B, ours achieves 87.0% Balanced Tool-Use Accuracy, outperforming Gold-SFT (66.3%), Oracle sequence distillation (82.3%), and off-policy token distillation (85.0%). The method scales consistently: an 8B teacher raises the same compact 1.7B student to 91.9%, while an 8B student reaches 93.0%. The resulting policies further transfer to clean histories, unseen functions, independently regenerated evaluation contexts, external tool-use benchmarks, and noisy multi-hop question answering. These results establish history reliability as a distinct tool-use bottleneck and demonstrate reliable-state policy transfer as an effective and scalable solution.

cs.AI↗

arXiv2026

WikiKV: Schema-Evolving Path-Indexed Storage for Hierarchical Knowledge Navigation

LLM-curated hierarchical knowledge bases, namely a tree-structured wiki whose nodes summarize an underlying corpus, have become a dominant substrate for retrieval-augmented applications, yet their storage layer is still treated as an implementation detail. This workload is hierarchical, query-intensive, and continuously evolving, and no existing storage model natively captures all three properties at once. We present WikiKV, a path-indexed key-value storage model purpose-built for this workload, comprising three components: (i) a data-driven schema that bootstraps the hierarchy via Intent-Anchored Schema Induction and refines it through Continuous Evolution Operators; (ii) a consistency protocol for the path-indexed storage model that precludes partial-read observations under concurrent offline rewrites without read-path locking; and (iii) a budgeted navigation operator whose search-accelerated routing reduces the expected number of LLM-assisted descent steps from d to O(1) while preserving anytime semantics with progressively refined answers. We evaluate WikiKV through real-world deployment for the WeChat Official Account AI Assistant and benchmark it against diverse baselines on the AuthTrace dataset, where it achieves balanced low per-operator latency across four query operators against relational, graph, and FS backends, and reaches 63.2% end-to-end answer correctness, exceeding multiple RAG baselines, with the gap widening on low- and high-fan-in multi-document questions. Ablation study further confirms the effectiveness of WikiKV's components.

cs.DB↗

arXiv2026

Semantic Flow Regularization: Teaching LLMs to Generate Diverse Yet Coherent Responses

When large language models are fine-tuned to generate persona- or tone-conditioned responses, their output diversity is severely limited--a failure we term Cross-Style Collapse. We trace this collapse to the cross-entropy objective, which under shared representations tends to suppress diverse continuations. We propose Semantic Flow Regularization (SFR), a lightweight auxiliary objective that supervises the backbone with continuous sentence-encoder embeddings of future segments via conditional flow matching. The stochastic flow source preserves multi-modality by construction; the flow-matching head is discarded at inference, adding zero deployment cost. On a large-scale industrial dialogue dataset (Qwen3-32B, 9 personas), SFR improves output diversity, style fidelity, and response quality over SFT. We further validate on the public LiveCodeBench-v5 (Qwen2.5-Coder-7B-Instruct), where SFR consistently improves pass@k, confirming generality beyond stylized dialogue. A controlled comparison on MBPP reveals Multi-Token Prediction to be a degenerate special case of SFR.

cs.CL↗

arXiv2026

MemCog: From Memory-as-Tool to Memory-as-Cognition in Conversational Agents

Existing agent memory systems universally follow what we term a Memory-as-Tool paradigm where a single query triggers one-shot retrieval of flat passage lists, suffering from passive invocation, reasoning-retrieval decoupling, and structural mismatch between retrieved fragments and the agent's navigational needs. We propose MemCog, a Memory-as-Cognition system that makes memory access an integral part of the reasoning process. MemCog organizes user knowledge as Navigable Memory Store with associative link graphs, exposes Cross-Dimensional Navigation Interface for multi-step reasoning-driven traversal, and employs Proactive Reasoning Protocol that drives agents to spontaneously initiate memory exploration from conversational context. We additionally construct ProactiveMemBench, the first benchmark for evaluating proactive memory triggering. Experiments show that MemCog achieves state-of-the-art on passive QA benchmarks (92.98 on LoCoMo, 95.8 on LongMemEval) while substantially outperforming baselines on ProactiveMemBench, demonstrating the advantage of Memory-as-Cognition.

cs.AI↗

arXiv2026

PARM: Pipeline-Adapted Reward Model

Reward models (RMs) are central to aligning large language models (LLMs) with human preferences, powering RLHF and advanced decoding strategies. While most prior work focuses on single-step generation, real-world applications increasingly adopt multi-stage LLM pipelines, where effective reward guidance remains underexplored. We investigate this through code generation for combinatorial optimization, constructing a pipeline that integrates reward models into both formulation and solution stages. We identify a critical challenge: inconsistency between reward model predictions and actual pipeline execution outcomes. To address this, we propose the Pipeline-Adapted Reward Model (PARM), which leverages pipeline-specific data and direct preference optimization to align rewards with downstream feedback. We instantiate PARM as a two-stage pipeline (formulation -> code generation) and evaluate it on four public optimization benchmarks, measuring execution rate and solving accuracy against baselines and sampling methods. A supplementary cross-domain experiment on GSM8K assesses transferability. Results demonstrate that PARM consistently improves pipeline output quality and stability, providing new insights into reward modeling for multi-stage LLM reasoning.

cs.AI↗

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

HHEML: Hybrid Homomorphic Encryption for Privacy-Preserving Machine Learning on Edge

Privacy-preserving machine learning (PPML) is an emerging topic to handle secure machine learning inference over sensitive data in untrusted environments. Fully homomorphic encryption (FHE) enables computation directly on encrypted data on the server side, making it a promising approach for PPML. However, it introduces significant communication and computation overhead on the client side, making it impractical for edge devices. Hybrid homomorphic encryption (HHE) addresses this limitation by combining symmetric encryption (SE) with FHE to reduce the computational cost on the client side, and combining with an FHE-friendly SE can also lessen the processing overhead on the server side, making it a more balanced and efficient alternative. Our work proposes a hardware-accelerated HHE architecture built around a lightweight symmetric cipher optimized for FHE compatibility and implemented as a dedicated hardware accelerator. To the best of our knowledge, this is the first design to integrate an end-to-end HHE framework with hardware acceleration. Beyond this, we also present several microarchitectural optimizations to achieve higher performance and energy efficiency. The proposed work is integrated into a full PPML pipeline, enabling secure inference with significantly lower latency and power consumption than software implementations. Our contributions validate the feasibility of low-power, hardware- accelerated HHE for edge deployment and provide a hardware- software co-design methodology for building scalable, secure machine learning systems in resource-constrained environments. Experiments on a PYNQ-Z2 platform with the MNIST dataset show over a 50x reduction in client-side encryption latency and nearly a 2x gain in hardware throughput compared to existing FPGA-based HHE accelerators.

cs.CR↗

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

IndusGCC: A Data Benchmark and Evaluation Framework for GUI-Based General Computer Control in Industrial Automation

As Industry 4.0 progresses, flexible manufacturing has become a cornerstone of modern industrial systems, with equipment automation playing a pivotal role. However, existing control software for industrial equipment, typically reliant on graphical user interfaces (GUIs) that require human interactions such as mouse clicks or screen touches, poses significant barriers to the adoption of code-based equipment automation. Recently, Large Language Model-based General Computer Control (LLM-GCC) has emerged as a promising approach to automate GUI-based operations. However, industrial settings pose unique challenges, including visually diverse, domain-specific interfaces and mission-critical tasks demanding high precision. This paper introduces IndusGCC, the first dataset and benchmark tailored to LLM-GCC in industrial environments, encompassing 448 real-world tasks across seven domains, from robotic arm control to production line configuration. IndusGCC features multimodal human interaction data with the equipment software, providing robust supervision for GUI-level code generation. Additionally, we propose a novel evaluation framework with functional and structural metrics to assess LLM-generated control scripts. Experimental results on mainstream LLMs demonstrate both the potential of LLM-GCC and the challenges it faces, establishing a strong foundation for future research toward fully automated factories. Our data and code are publicly available at: \href{https://github.com/Golden-Arc/IndustrialLLM}{https://github.com/Golden-Arc/IndustrialLLM.

eess.SY↗

arXiv2021

HMD-AMP: Protein Language-Powered Hierarchical Multi-label Deep Forest for Annotating Antimicrobial Peptides

Identifying the targets of an antimicrobial peptide is a fundamental step in studying the innate immune response and combating antibiotic resistance, and more broadly, precision medicine and public health. There have been extensive studies on the statistical and computational approaches to identify (i) whether a peptide is an antimicrobial peptide (AMP) or a non-AMP and (ii) which targets are these sequences effective to (Gram-positive, Gram-negative, etc.). Despite the existing deep learning methods on this problem, most of them are unable to handle the small AMP classes (anti-insect, anti-parasite, etc.). And more importantly, some AMPs can have multiple targets, which the previous methods fail to consider. In this study, we build a diverse and comprehensive multi-label protein sequence database by collecting and cleaning amino acids from various AMP databases. To generate efficient representations and features for the small classes dataset, we take advantage of a protein language model trained on 250 million protein sequences. Based on that, we develop an end-to-end hierarchical multi-label deep forest framework, HMD-AMP, to annotate AMP comprehensively. After identifying an AMP, it further predicts what targets the AMP can effectively kill from eleven available classes. Extensive experiments suggest that our framework outperforms state-of-the-art models in both the binary classification task and the multi-label classification task, especially on the minor classes.The model is robust against reduced features and small perturbations and produces promising results. We believe HMD-AMP contributes to both the future wet-lab investigations of the innate structural properties of different antimicrobial peptides and build promising empirical underpinnings for precise medicine with antibiotics.

cs.LG↗

arXiv2018

MRI Cross-Modality NeuroImage-to-NeuroImage Translation

We present a cross-modality generation framework that learns to generate translated modalities from given modalities in MR images without real acquisition. Our proposed method performs NeuroImage-to-NeuroImage translation (abbreviated as N2N) by means of a deep learning model that leverages conditional generative adversarial networks (cGANs). Our framework jointly exploits the low-level features (pixel-wise information) and high-level representations (e.g. brain tumors, brain structure like gray matter, etc.) between cross modalities which are important for resolving the challenging complexity in brain structures. Our framework can serve as an auxiliary method in clinical diagnosis and has great application potential. Based on our proposed framework, we first propose a method for cross-modality registration by fusing the deformation fields to adopt the cross-modality information from translated modalities. Second, we propose an approach for MRI segmentation, translated multichannel segmentation (TMS), where given modalities, along with translated modalities, are segmented by fully convolutional networks (FCN) in a multichannel manner. Both of these two methods successfully adopt the cross-modality information to improve the performance without adding any extra data. Experiments demonstrate that our proposed framework advances the state-of-the-art on five brain MRI datasets. We also observe encouraging results in cross-modality registration and segmentation on some widely adopted brain datasets. Overall, our work can serve as an auxiliary method in clinical diagnosis and be applied to various tasks in medical fields. Keywords: image-to-image, cross-modality, registration, segmentation, brain MRI

cs.CV↗