arXiv subjects

Xuan Feng

Publications and source records attributed to Xuan Feng.

At least 19 recordsLinked to original sources

arXiv2026

PlanPO: Group Planning-Aware Policy Optimization for Multi-Turn Agentic LLMs

Group-relative policy optimization has emerged as a key paradigm for training agentic large language models (LLMs) on multi-turn interactive tasks. However, most existing variants fail to distinguish advantages among successful trajectories even when these trajectories differ substantially in their interaction efficiency. For instance, circuitous successes are often assigned the identical outcome reward, causing advantage collapse and severe performance bottlenecks. To this end, we propose Group Planning-aware Policy Optimization (PlanPO), a simple yet effective RL method for learning generalizable planning abilities beyond task-specific high-quality behavior patterns. Specifically, PlanPO introduces coarse-to-fine advantage signals, which capture the relative differences in trajectory-level lengths and turn-level response lengths conditioned on successful trajectories sampled for the same task. Within the group-relative optimization structure, this enables agents to actively learn generalizable and deliberate behaviors spanning interaction planning and textual generation from high-quality rollouts, without degenerating into vanilla length minimization. Experimentally, PlanPO improves over GRPO by 27.2\% on average across the challenging multi-turn benchmarks ALFWorld, WebShop, and SciWorld, outperforming recent powerful baselines while incurring negligible additional training cost.

cs.AI↗

arXiv2026

Where Detectors Fail: Closing the Tail-Domain Gap with Expert-Guided Mutual Distillation

Multimodal fake news detectors often generalize poorly across domains because they learn to trust unreliable evidence: domain-specific shortcuts amplified by imbalanced data and semantically inconsistent text-image pairs that make cross-modal evidence unreliable. We propose Expert-Guided Mutual Distillation (EGMD), which learns what evidence to trust across the prediction pipeline. At the input level, input-level calibration encodes pair-level coherence as a shared gain before fusion. At the representation level, an expert-guided teacher aligns domain statistics and encourages domain-specific patterns to concentrate in specialized experts. At the decision level, prototype-anchored domain-specific students use mutual learning and dual-channel distillation to inherit the teacher's feature geometry and calibrated predictions while discouraging local domain priors. We further construct Weibo_Balanced, a domain-balanced benchmark that isolates the effect of imbalance on generalization. Across four datasets in two languages, EGMD achieves state-of-the-art accuracy while reducing domain bias by up to 57.3%.

cs.CL↗

arXiv2026

From Noisy Traces to Root Causes: Structural Trajectory Analysis and Causal Extraction for Agent Optimization

The optimization of long-horizon agents increasingly relies on reflection-based mechanisms, where a large language model (LLM) acts as an optimizer to diagnose agent failures and improve agent policies. However, real execution traces are difficult to use directly for optimization: large trace collections are often redundant and heterogeneous, making optimization inefficient and prone to overfitting to low-value failures; meanwhile, each individual trajectory also contains many irrelevant steps, while naive context reduction methods such as truncation or sliding windows can discard causally important evidence and produce misleading optimization signals. To resolve this dilemma, we introduce STRACE (Structural TRajectory Analysis and Causal Extraction), a framework that constructs high signal-noise optimization contexts for more precise and effective optimization. At the batch level, STRACE mines failure patterns to filter redundant traces and retain representative failures; within each selected trace, it performs causal localization over a textual dependency graph to remove non-causal steps and identify the true root-cause module for optimization. Empirical results demonstrate that STRACE significantly outperforms standard context-filtering baselines. Notably, on a challenging formal verification task (VeruSAGE-Bench), it successfully optimizes human-expert designed agents, delivering $1.4\times$ success-rate improvement (42.5% to 58.5%). The code is available at https://github.com/moomight/STRACE .

cs.CL↗

arXiv2026

RustMizan: A Compilable, Contamination-Aware Benchmarking Framework for Rust Vulnerabilities

LLM agents are increasingly applied to vulnerability analysis, but existing benchmarks have not kept pace. They typically rely on small non-compilable snippets, focus on binary classification (vulnerable or not), and do not account for the risk that publicly-released datasets are part of model training corpora. We introduce RustMizan, a benchmarking framework for Rust vulnerability analysis that addresses these gaps. RustMizan contains compilable code variants at the crate, file, and function levels, with annotations for binary vulnerability detection, CWE classification, and function- and line-level localization. A paired mutation framework produces semantics-preserving code mutants for contamination testing and robustness probing. Across four frontier models in an agentic setup with command-line access, binary classification sits in the 56-65% range, but line localization F1 stays near 20%, and adversarial cues drop line F1 by about 27%.

cs.CR↗

arXiv2026

Self-Debias: Self-correcting for Debiasing Large Language Models

Although Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, inherent social biases often cascade throughout the Chain-of-Thought (CoT) process, leading to continuous "Bias Propagation". Existing debiasing methods primarily focus on static constraints or external interventions, failing to identify and interrupt this propagation once triggered. To address this limitation, we introduce Self-Debias, a progressive framework designed to instill intrinsic self-correction capabilities. Specifically, we reformulate the debiasing process as a strategic resource redistribution problem, treating the model's output probability mass as a limited resource to be reallocated from biased heuristics to unbiased reasoning paths. Unlike standard preference optimization which applies broad penalties, Self-Debias employs a fine-grained trajectory-level objective subject to dynamic debiasing constraints. This enables the model to selectively revise biased reasoning suffixes while preserving valid contextual prefixes. Furthermore, we integrate an online self-improvement mechanism utilizing consistency filtering to autonomously synthesize supervision signals. With merely 20k annotated samples, Self-Debias activates efficient self-correction, achieving superior debiasing performance while preserving general reasoning capabilities without continuous external oversight.

cs.CL↗

arXiv2026

SLARM: Streaming and Language-Aligned Reconstruction Model for Dynamic Scenes

We propose SLARM, a feed-forward model that unifies dynamic scene reconstruction, semantic understanding, and real-time streaming inference. SLARM captures complex, non-uniform motion through higher-order motion modeling, trained solely on differentiable renderings without any flow supervision. Besides, SLARM distills semantic features from LSeg to obtain language-aligned representations. This design enables semantic querying via natural language, and the tight coupling between semantics and geometry further enhances the accuracy and robustness of dynamic reconstruction. Moreover, SLARM processes image sequences using window-based causal attention, achieving stable, low-latency streaming inference without accumulating memory cost. Within this unified framework, SLARM achieves state-of-the-art results in dynamic estimation, rendering quality, and scene parsing, improving motion accuracy by 21%, reconstruction PSNR by 1.6 dB, and segmentation mIoU by 20% over existing methods.

cs.CV↗

arXiv2026

CyberThreat-Eval: Can Large Language Models Automate Real-World Threat Research?

Analyzing Open Source Intelligence (OSINT) from large volumes of data is critical for drafting and publishing comprehensive CTI reports. This process usually follows a three-stage workflow -- triage, deep search and TI drafting. While Large Language Models (LLMs) offer a promising route toward automation, existing benchmarks still have limitations. These benchmarks often consist of tasks that do not reflect real-world analyst workflows. For example, human analysts rarely receive tasks in the form of multiple-choice questions. Also, existing benchmarks often rely on model-centric metrics that emphasize lexical overlap rather than actionable, detailed insights essential for security analysts. Moreover, they typically fail to cover the complete three-stage workflow. To address these issues, we introduce CyberThreat-Eval, which is collected from the daily CTI workflow of a world-leading company. This expert-annotated benchmark assesses LLMs on practical tasks across all three stages as mentioned above. It utilizes analyst-centric metrics that measure factual accuracy, content quality, and operational costs. Our evaluation using this benchmark reveals important insights into the limitations of current LLMs. For example, LLMs often lack the nuanced expertise required to handle complex details and struggle to distinguish between correct and incorrect information. To address these challenges, the CTI workflow incorporates both external ground-truth databases and human expert knowledge. TRA allows human experts to iteratively provide feedback for continuous improvement. The code is available at \href{https://github.com/xschen-beb/CyberThreat-Eval}{\texttt{GitHub}} and \href{https://huggingface.co/datasets/xse/CyberThreat-Eval}{\texttt{HuggingFace}}.

cs.CR↗

arXiv2026

Lightweight Yet Secure: Secure Scripting Language Generation via Lightweight LLMs

The security of scripting languages such as PowerShell is critical given their powerful automation and administration capabilities, often exercised with elevated privileges. Today, securing these languages still demands substantial human effort to craft and enforce rules, imposing heavy burdens on typical administrators and creating critical production risks (e.g., misoperations that shut down servers).Large language models (LLMs) have demonstrated strong capabilities in code generation, vulnerability detection, and automated repair for languages like Python and JavaScript. However, their ability to assist with generating secure scripting-language code remains largely underexplored. In this paper, we present SecGenEval-PS, a benchmark designed to systematically evaluate LLMs on secure scripting generation, security analysis, and automated repair. Our results show that both proprietary and open-source models fall short in these areas. For instance, over 60% of PowerShell scripts produced by GPT-4o and o3-mini are insecure without structured guidance.To bridge this gap, we propose PSSec, a framework that combines data synthesis with fine-tuning to enhance model security capabilities. We develop a self-debugging agent that integrates static analyzers with the reasoning abilities of advanced LLMs to synthesize large-scale structured triplets of insecure scripts, violation analyses, and corresponding repairs. We then fine-tune lightweight LLMs (as small as 1.7B parameters) using supervised fine-tuning (SFT) and reinforcement learning (RL), enabling security-aware reasoning and the generation of secure PowerShell code.Across multiple LLM families, including GPT and Qwen, \textit{PSSec}-trained models match or surpass general-purpose large models on PowerShell security tasks while reducing inference cost by more than an order of magnitude.

cs.CR↗

arXiv2025

C2PO: Diagnosing and Disentangling Bias Shortcuts in LLMs

Bias in Large Language Models (LLMs) poses significant risks to trustworthiness, manifesting primarily as stereotypical biases (e.g., gender or racial stereotypes) and structural biases (e.g., lexical overlap or position preferences). However, prior paradigms typically address these in isolation, often mitigating one at the expense of exacerbating the other. To address this, we conduct a systematic exploration of these reasoning failures and identify a primary inducement: the latent spurious feature correlations within the input that drive these erroneous reasoning shortcuts. Driven by these findings, we introduce Causal-Contrastive Preference Optimization (C2PO), a unified alignment framework designed to tackle these specific failures by simultaneously discovering and suppressing these correlations directly within the optimization process. Specifically, C2PO leverages causal counterfactual signals to isolate bias-inducing features from valid reasoning paths, and employs a fairness-sensitive preference update mechanism to dynamically evaluate logit-level contributions and suppress shortcut features. Extensive experiments across multiple benchmarks covering stereotypical bias (BBQ, Unqover), structural bias (MNLI, HANS, Chatbot, MT-Bench), out-of-domain fairness (StereoSet, WinoBias), and general utility (MMLU, GSM8K) demonstrate that C2PO effectively mitigates stereotypical and structural biases while preserving robust general reasoning capabilities.

cs.CL↗

arXiv2025

Fast ground penetrating radar dual-parameter full waveform inversion method accelerated by hybrid compilation of CUDA kernel function and PyTorch

This study proposes a high-performance dual-parameter full waveform inversion framework (FWI) for ground-penetrating radar (GPR), accelerated through the hybrid compilation of CUDA kernel functions and PyTorch. The method leverages the computational efficiency of GPU programming while preserving the flexibility and usability of Python-based deep learning frameworks. By integrating customized CUDA kernels into PyTorch's automatic differentiation mechanism, the framework enables accurate and efficient inversion of both dielectric permittivity and electrical conductivity. Experimental evaluations on synthetic data and real wavefield data demonstrate that the proposed method achieves dual-parameter FWI for GPR data while maintaining high accuracy. Moreover, the framework is flexible and extensible, supporting optional regularization strategies such as total variation and multi-scale inversion. These features make the proposed approach a practical and scalable framework for rapid GPR-based subsurface imaging in applications including civil engineering, environmental monitoring, and geophysical exploration.

physics.geo-ph↗

arXiv2025

Unlocking the Capabilities of Large Vision-Language Models for Generalizable and Explainable Deepfake Detection

Current Large Vision-Language Models (LVLMs) have demonstrated remarkable capabilities in understanding multimodal data, but their potential remains underexplored for deepfake detection due to the misalignment of their knowledge and forensics patterns. To this end, we present a novel framework that unlocks LVLMs' potential capabilities for deepfake detection. Our framework includes a Knowledge-guided Forgery Detector (KFD), a Forgery Prompt Learner (FPL), and a Large Language Model (LLM). The KFD is used to calculate correlations between image features and pristine/deepfake image description embeddings, enabling forgery classification and localization. The outputs of the KFD are subsequently processed by the Forgery Prompt Learner to construct fine-grained forgery prompt embeddings. These embeddings, along with visual and question prompt embeddings, are fed into the LLM to generate textual detection responses. Extensive experiments on multiple benchmarks, including FF++, CDF2, DFD, DFDCP, DFDC, and DF40, demonstrate that our scheme surpasses state-of-the-art methods in generalization performance, while also supporting multi-turn dialogue capabilities.

cs.CV↗

arXiv2025

Beyond Prompt Content: Enhancing LLM Performance via Content-Format Integrated Prompt Optimization

Large Language Models (LLMs) have shown significant capability across various tasks, with their real-world effectiveness often driven by prompt design. While recent research has focused on optimizing prompt content, the role of prompt formatting, a critical but often overlooked dimension, has received limited systematic investigation. In this paper, we introduce Content-Format Integrated Prompt Optimization (CFPO), an innovative methodology that jointly optimizes both prompt content and formatting through an iterative refinement process. CFPO leverages natural language mutations to explore content variations and employs a dynamic format exploration strategy that systematically evaluates diverse format options. Our extensive evaluations across multiple tasks and open-source LLMs demonstrate that CFPO demonstrates measurable performance improvements compared to content-only optimization methods. This highlights the importance of integrated content-format optimization and offers a practical, model-agnostic approach to enhancing LLM performance. Code is available at https://github.com/HenryLau7/CFPO.

cs.CL↗

arXiv2025

Sigma: Differential Rescaling of Query, Key and Value for Efficient Language Models

We introduce Sigma, an efficient large language model specialized for the system domain, empowered by a novel architecture including DiffQKV attention, and pre-trained on our meticulously collected system domain data. DiffQKV attention significantly enhances the inference efficiency of Sigma by optimizing the Query (Q), Key (K), and Value (V) components in the attention mechanism differentially, based on their varying impacts on the model performance and efficiency indicators. Specifically, we (1) conduct extensive experiments that demonstrate the model's varying sensitivity to the compression of K and V components, leading to the development of differentially compressed KV, and (2) propose augmented Q to expand the Q head dimension, which enhances the model's representation capacity with minimal impacts on the inference speed. Rigorous theoretical and empirical analyses reveal that DiffQKV attention significantly enhances efficiency, achieving up to a 33.36% improvement in inference speed over the conventional grouped-query attention (GQA) in long-context scenarios. We pre-train Sigma on 6T tokens from various sources, including 19.5B system domain data that we carefully collect and 1T tokens of synthesized and rewritten data. In general domains, Sigma achieves comparable performance to other state-of-arts models. In the system domain, we introduce the first comprehensive benchmark AIMicius, where Sigma demonstrates remarkable performance across all tasks, significantly outperforming GPT-4 with an absolute improvement up to 52.5%.

cs.CL↗

arXiv2024

Learning from Mistakes: Self-correct Adversarial Training for Chinese Unnatural Text Correction

Unnatural text correction aims to automatically detect and correct spelling errors or adversarial perturbation errors in sentences. Existing methods typically rely on fine-tuning or adversarial training to correct errors, which have achieved significant success. However, these methods exhibit poor generalization performance due to the difference in data distribution between training data and real-world scenarios, known as the exposure bias problem. In this paper, we propose a self-correct adversarial training framework for \textbf{L}earn\textbf{I}ng from \textbf{MI}s\textbf{T}akes (\textbf{LIMIT}), which is a task- and model-independent framework to correct unnatural errors or mistakes. Specifically, we fully utilize errors generated by the model that are actively exposed during the inference phase, i.e., predictions that are inconsistent with the target. This training method not only simulates potential errors in real application scenarios, but also mitigates the exposure bias of the traditional training process. Meanwhile, we design a novel decoding intervention strategy to maintain semantic consistency. Extensive experimental results on Chinese unnatural text error correction datasets show that our proposed method can correct multiple forms of errors and outperforms the state-of-the-art text correction methods. In addition, extensive results on Chinese and English datasets validate that LIMIT can serve as a plug-and-play defense module and can extend to new models and datasets without further training.

cs.CL↗

arXiv2024

Glass Transition in Monolayers of Rough Colloidal Ellipsoids

Structure-dynamics correlation is one of the major ongoing debates in the glass transition, although a number of structural features have been found connected to the dynamic heterogeneity in different glass-forming colloidal systems. Here using colloidal experiments combined with coarse-grained molecular dynamics simulations, we investigate the glass transition in monolayers of rough colloidal ellipsoids. Compared with smooth colloidal ellipsoids, the surface roughness of ellipsoids is found to significantly change the nature of glass transition. In particular, we find that the surface roughness induced by coating only a few small hemispheres on the ellipsoids can eliminate the existence of orientational glass and the two-step glass transition found in monolayers of smooth ellipsoids. This is due to the surface roughness-induced coupling between the translational and rotational degrees of freedom in colloidal ellipsoids, which also destroys the structure-dynamics correlation found in glass-forming suspensions of colloidal ellipsoids. Our results not only suggest a new way of using surface roughness to manipulate the glass transition in colloidal systems, but also highlight the importance of detailed particle shape on the glass transition and structure-dynamics correlation in suspensions of anisotropic colloids.

cond-mat.soft↗

arXiv2024

Enabling Privacy-Preserving Cyber Threat Detection with Federated Learning

Despite achieving good performance and wide adoption, machine learning based security detection models (e.g., malware classifiers) are subject to concept drift and evasive evolution of attackers, which renders up-to-date threat data as a necessity. However, due to enforcement of various privacy protection regulations (e.g., GDPR), it is becoming increasingly challenging or even prohibitive for security vendors to collect individual-relevant and privacy-sensitive threat datasets, e.g., SMS spam/non-spam messages from mobile devices. To address such obstacles, this study systematically profiles the (in)feasibility of federated learning for privacy-preserving cyber threat detection in terms of effectiveness, byzantine resilience, and efficiency. This is made possible by the build-up of multiple threat datasets and threat detection models, and more importantly, the design of realistic and security-specific experiments. We evaluate FL on two representative threat detection tasks, namely SMS spam detection and Android malware detection. It shows that FL-trained detection models can achieve a performance that is comparable to centrally trained counterparts. Also, most non-IID data distributions have either minor or negligible impact on the model performance, while a label-based non-IID distribution of a high extent can incur non-negligible fluctuation and delay in FL training. Then, under a realistic threat model, FL turns out to be adversary-resistant to attacks of both data poisoning and model poisoning. Particularly, the attacking impact of a practical data poisoning attack is no more than 0.14\% loss in model accuracy. Regarding FL efficiency, a bootstrapping strategy turns out to be effective to mitigate the training delay as observed in label-based non-IID scenarios.

cs.CR↗

arXiv2024

Port Forwarding Services Are Forwarding Security Risks

We conduct the first comprehensive security study on representative port forwarding services (PFS), which emerge in recent years and make the web services deployed in internal networks available on the Internet along with better usability but less complexity compared to traditional techniques (e.g., NAT traversal techniques). Our study is made possible through a set of novel methodologies, which are designed to uncover the technical mechanisms of PFS, experiment attack scenarios for PFS protocols, automatically discover and snapshot port-forwarded websites (PFWs) at scale, and classify PFWs into well-observed categories. Leveraging these methodologies, we have observed the widespread adoption of PFS with millions of PFWs distributed across tens of thousands of ISPs worldwide. Furthermore, 32.31% PFWs have been classified into website categories that serve access to critical data or infrastructure, such as, web consoles for industrial control systems, IoT controllers, code repositories, and office automation systems. And 18.57% PFWs didn't enforce any access control for external visitors. Also identified are two types of attacks inherent in the protocols of Oray (one well-adopted PFS provider), and the notable abuse of PFSes by malicious actors in activities such as malware distribution, botnet operation and phishing.

cs.CR↗

arXiv2023

Dual-Teacher De-biasing Distillation Framework for Multi-domain Fake News Detection

Multi-domain fake news detection aims to identify whether various news from different domains is real or fake and has become urgent and important. However, existing methods are dedicated to improving the overall performance of fake news detection, ignoring the fact that unbalanced data leads to disparate treatment for different domains, i.e., the domain bias problem. To solve this problem, we propose the Dual-Teacher De-biasing Distillation framework (DTDBD) to mitigate bias across different domains. Following the knowledge distillation methods, DTDBD adopts a teacher-student structure, where pre-trained large teachers instruct a student model. In particular, the DTDBD consists of an unbiased teacher and a clean teacher that jointly guide the student model in mitigating domain bias and maintaining performance. For the unbiased teacher, we introduce an adversarial de-biasing distillation loss to instruct the student model in learning unbiased domain knowledge. For the clean teacher, we design domain knowledge distillation loss, which effectively incentivizes the student model to focus on representing domain features while maintaining performance. Moreover, we present a momentum-based dynamic adjustment algorithm to trade off the effects of two teachers. Extensive experiments on Chinese and English datasets show that the proposed method substantially outperforms the state-of-the-art baseline methods in terms of bias metrics while guaranteeing competitive performance.

cs.CL↗