arXiv subjects

Yiran Zhang

Publications and source records attributed to Yiran Zhang.

At least 37 records · Page 2Linked to original sources

arXiv2026

Securing Code Understanding: Detecting Natural Backdoor Vulnerability in Code Language Models

Code Language Models (CodeLMs) have become integral to software engineering, significantly advancing code intelligence tasks. However, their widespread adoption has raised critical security concerns, particularly regarding susceptibility to backdoor attacks. Recent studies have uncovered naturally occurring backdoors, referred to as natural backdoors, in normally trained deep learning models. Despite posing threats as serious as those introduced through data poisoning, security implications of natural backdoor vulnerabilities in CodeLMs remain poorly understood. In this paper, we conduct a thorough empirical study of natural backdoor vulnerabilities in CodeLMs across various model architectures and code intelligence tasks. Specifically, we examine potential natural backdoor vulnerabilities across 44 scenarios, demonstrating that natural backdoors are prevalent and intrinsic to CodeLMs. We reveal differences between injected and natural backdoor vulnerabilities at both the model and parameter levels. We then analyze the transferability of natural backdoor vulnerabilities from three perspectives: datasets, model architectures, and shared knowledge. We further investigate the causes of natural backdoors from two aspects: training datasets and the model training procedure. We evaluate existing backdoor defense techniques, including pre-training, in-training, and post-training defenses, in mitigating natural backdoors. Finally, we propose ScanNBT, a novel detection method designed to improve comprehensive detection of natural backdoor vulnerabilities in CodeLMs. We aim for our findings to enhance understanding of these vulnerabilities and provide insights for strengthening CodeLM security against backdoor threats.

cs.CR↗

arXiv2026

Bridging Requirements and Architecture: Multi-Agent Orchestration with External Knowledge and Hierarchical Memory

Software architecture design is a critical yet inherently complex and knowledge-intensive phase that requires balancing competing quality attributes and adapting to evolving requirements. Traditionally, this process has been time-consuming, labor-intensive, and heavily reliant on architects, often resulting in limited exploration of alternative architectural decompositions and styles, especially under the pressures of agile development. While LLM-based agents have shown promising performance across various software engineering tasks, their application to architecture design remains relatively scarce and requires systematic exploration. To address these challenges, we proposed MAAD (Multi-Agent Architecture Design), a knowledge-driven framework that orchestrates four specialized agents (i.e., Analyst, Modeler, Designer and Evaluator) to autonomously and collaboratively transform requirements specifications into comprehensive, multi-view architectural blueprints with quality attribute assessments. MAAD incorporates RAG to inject recognized architectural standards and patterns into the workflow and leverages a hierarchical memory mechanism that captures design history for iterative refinement. We evaluated MAAD through comparative experiments against MetaGPT, using quantitative architecture-level metrics across 10 case studies and qualitative feedback from industry architects on 10 real-world specifications. Results show that MAAD generates more complete, modular, and traceable architectures than the baseline, and its dedicated Evaluator agent autonomously produces structured quality evaluation reports that significantly reduce manual validation efforts. Furthermore, we found that the quality of the generated architecture heavily depends on the underlying LLM's reasoning capacity, with GPT-5.2 and Qwen3.5 outperforming other models across most evaluation settings.

cs.SE↗

arXiv2026

RLCracker: Evaluating the Worst-Case Vulnerability of LLM Watermarks with Adaptive RL Attacks

Large language model (LLM) watermarking has shown promise in detecting AI-generated content and mitigating misuse, with prior work claiming robustness against paraphrasing and text editing. In this paper, we argue that existing evaluations are not sufficiently adversarial, obscuring critical vulnerabilities and overstating the security. To address this, we introduce the adaptive robustness radius, a formal metric that quantifies the worst-case resilience of watermarks against adaptive adversaries. By lifting the paraphrase space into a KL-divergence ball, we approximate this radius and theoretically demonstrate that optimizing the attack context and model parameters can significantly reduce the approximate radius, making watermarks highly vulnerable to paraphrase attacks. Leveraging this insight, we propose RLCracker, a reinforcement learning (RL)-based adaptive attack that erases watermark signals with limited watermarked examples and limited access to the detector. Despite weak supervision, it empowers a 3B model to achieve 98.5% removal success with minimal semantic shift on 1,500-token Unigram-marked texts after training on only 100 short samples. This performance dramatically exceeds 6.75% by GPT-4o and generalizes across five model sizes over ten watermarking schemes. Our code is available at https://github.com/OTT0-OTO/RLCracker.

cs.CR↗

arXiv2026

Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning

Long chain-of-thought (CoT) reasoning improves large vision--language models, but visual information often fades during generation, limiting long-horizon multimodal reasoning. Existing methods either re-inject vision at inference or train policies for stronger grounding, but where to intervene relies on perception heuristics rather than principled gain analysis, and how local visual influence propagates remains implicit. We study this problem from an information-theoretic standpoint and derive a lower bound on the downstream visual gain of a one-step intervention, which suggests two factors: local branching room (token entropy) and downstream visual propagation potential (suffix divergence from a vision-marginalized reference). Guided by this analysis, we propose reflection-anchor policy optimization (RAPO), a GRPO-based policy optimization method that selects high-entropy reflection anchors and optimizes a chain-masked finite-window KL surrogate for downstream visual dependence. Experiments on reasoning-intensive and general-domain benchmarks show that RAPO delivers substantial gains over strong baselines across multiple LVLM backbones. Mechanism analyses further indicate that reflection anchors are enriched for visually sensitive decision points and that RAPO increases contrastive visual-dependence signals along generated trajectories.

cs.CV↗

arXiv2026

Bridging the Gap between User Intent and LLM: A Requirement Alignment Approach for Code Generation

Code generation refers to automatically producing executable programs from user requirements. Recently, researchers have explored approaches to enhance the correctness of generated code with advanced large language models. Although achieving improvements, existing approaches focus on designing reasoning strategies or post-refinement methods to enhance code generation performance. Despite their differences, all these methods share a common assumption: the LLM can correctly understand the given requirement. However, this assumption does not always hold. To fill this gap, we propose REA-Coder, a requirement alignment approach to enhance the code generation performance of LLMs. REA-Coder involves first identifying the requirement content that does not align with LLMs and aligning the requirements. Then, based on the aligned requirements, LLMs generate code and further verify whether the generated code aligns with the requirements, iterating this process of requirement alignment and code generation until generating correct code or achieving the maximum number of iterations. Experimental results show that REA-Coder outperforms all advanced baselines on four LLMs across five programming benchmarks. Concretely, REA-Coder achieves average improvements of 7.93%, 30.25%, 26.75%, 8.59%, and 8.64% on the five benchmark datasets, demonstrating the effectiveness of requirement alignment for improving the code generation performance of LLMs.

cs.SE↗

arXiv2026

Cost-optimal Sequential Testing via Doubly Robust Q-learning

Clinical decision-making often involves selecting tests that are costly, invasive, or time-consuming, motivating individualized, sequential strategies for what to measure and when to stop ascertaining. We study the problem of learning cost-optimal sequential decision policies from retrospective data, where test availability depends on prior results, inducing informative missingness. Under a sequential missing-at-random mechanism, we develop a doubly robust Q-learning framework for estimating optimal policies. The method introduces path-specific inverse probability weights that account for heterogeneous test trajectories and satisfy a normalization property conditional on the observed history. By combining these weights with auxiliary contrast models, we construct orthogonal pseudo-outcomes that enable unbiased policy learning when either the acquisition model or the contrast model is correctly specified. We establish oracle inequalities for the stage-wise contrast estimators, along with convergence rates, regret bounds, and misclassification rates for the learned policy. Simulations demonstrate improved cost-adjusted performance over weighted and complete-case baselines, and an application to a prostate cancer cohort study illustrates how the method reduces testing cost without compromising predictive accuracy.

stat.ML↗

arXiv2026

RLSpoofer: A Lightweight Evaluator for LLM Watermark Spoofing Resilience

Large language model (LLM) watermarking has emerged as a promising approach for detecting and attributing AI-generated text, yet its robustness to black-box spoofing remains insufficiently evaluated. Existing evaluation methods often demand extensive datasets and white-box access to algorithmic internals, limiting their practical applicability. In this paper, we study watermark resilience against spoofing fundamentally from a distributional perspective. We first establish a \textit{local capacity bottleneck}, which theoretically characterizes the probability mass that can be reallocated under KL-bounded local updates while preserving semantic fidelity. Building on this, we propose RLSpoofer, a reinforcement learning-based black-box spoofing attack that requires only 100 human-watermarked paraphrase training pairs and zero access to the watermarking internals or detectors. Despite weak supervision, it empowers a 4B model to achieve a 62.0\% spoof success rate with minimal semantic shift on PF-marked texts, dwarfing the 6\% of baseline models trained on up to 10,000 samples. Our findings expose the fragile spoofing resistance of current LLM watermarking paradigms, providing a lightweight evaluation framework and stressing the urgent need for more robust schemes.

cs.CR↗

arXiv2026

From Context to Intent: Reasoning-Guided Function-Level Code Completion

The growing capabilities of Large Language Models (LLMs) have led to their widespread adoption for function completion within code repositories. Recent studies on such tasks show promising results when explicit instructions, often in the form of docstrings, are available to guide the completion. However, in real-world scenarios, clear docstrings are frequently absent. Under such conditions, LLMs typically fail to produce accurate completions. To enable more automated and accurate function completion in such settings, we aim to enable LLMs to accurately infer the developer's intent prior to code completion. Our key insight is that the preceding code, namely the code context before the function to be completed, often contains valuable cues that help the model understand the intended functionality. However, inferring intent from such implicit context is non-trivial and constitutes a core challenge in function-level code completion. To tackle this challenge, inspired by how humans interpret context, we propose a reasoning-based prompting framework that guides LLMs to utilize these contextual cues to infer intent step by step. To incentivize LLMs to reason through the preceding code and infer intent, we further curate a dataset of 40k examples, each annotated with intermediate reasoning traces and corresponding docstrings. Extensive experiments on DevEval and ComplexCodeEval demonstrate consistent performance improvements across multiple models, achieving over 25% relative gains in pass@1 for both DeepSeekCoder and CodeLLaMA families. Building upon our framework, we further develop an intent-interactive platform that supports lightweight human feedback. This platform allows developers to select from a set of candidate intentions or edit the intent to better guide the model. Our experiments show that this interactive approach leads to further performance improvements.

cs.SE↗

arXiv2026

GeoChemAD: Benchmarking Unsupervised Geochemical Anomaly Detection for Mineral Exploration

Geochemical anomaly detection plays a critical role in mineral exploration as deviations from regional geochemical baselines may indicate mineralization. Existing studies suffer from two key limitations: (1) single region scenarios which limit model generalizability; (2) proprietary datasets, which makes result reproduction unattainable. In this work, we introduce \textbf{GeoChemAD}, an open-source benchmark dataset compiled from government-led geological surveys, covering multiple regions, sampling sources, and target elements. The dataset comprises eight subsets representing diverse spatial scales and sampling conditions. To establish strong baselines, we reproduce and benchmark a range of unsupervised anomaly detection methods, including statistical models, generative and transformer-based approaches. Furthermore, we propose \textbf{GeoChemFormer}, a transformer-based framework that leverages self-supervised pretraining to learn target-element-aware geochemical representations for spatial samples. Extensive experiments demonstrate that GeoChemFormer consistently achieves superior and robust performance across all eight subsets, outperforming existing unsupervised methods in both anomaly detection accuracy and generalization capability. The proposed dataset and framework provide a foundation for reproducible research and future development in this direction.

cs.LG↗

arXiv2026

POLAR: A Benchmark for Multilingual, Multicultural, and Multi-Event Online Polarization

Online polarization poses a growing challenge for democratic discourse, yet most computational social science research remains monolingual, culturally narrow, or event-specific. We introduce POLAR, a multilingual, multicultural, and multi-event dataset with over 110K instances in 22 languages drawn from diverse online platforms and real-world events. Polarization is annotated along three axes, namely detection, type, and manifestation, using a variety of annotation platforms adapted to each cultural context. We conduct two main experiments: (1) fine-tuning six pretrained small language models; and (2) evaluating a range of open and closed large language models in few-shot and zero-shot settings. The results show that, while most models perform well in binary polarization detection, they achieve substantially lower performance when predicting polarization types and manifestations. These findings highlight the complex, highly contextual nature of polarization and demonstrate the need for robust, adaptable approaches in NLP and computational social science. All resources will be released to support further research and effective mitigation of digital polarization globally.

cs.CL↗

arXiv2026

Cutting the Gordian Knot: Detecting Malicious PyPI Packages via a Knowledge-Mining Framework

The Python Package Index (PyPI) has become a target for malicious actors, yet existing detection tools generate false positive rates of 15-30%, incorrectly flagging one-third of legitimate packages as malicious. This problem arises because current tools rely on simple syntactic rules rather than semantic understanding, failing to distinguish between identical API calls serving legitimate versus malicious purposes. To address this challenge, we propose PyGuard, a knowledge-driven framework that converts detection failures into useful behavioral knowledge by extracting patterns from existing tools' false positives and negatives. Our method utilizes hierarchical pattern mining to identify behavioral sequences that distinguish malicious from benign code, employs Large Language Models to create semantic abstractions beyond syntactic variations, and combines this knowledge into a detection system that integrates exact pattern matching with contextual reasoning. PyGuard achieves 99.50% accuracy with only 2 false positives versus 1,927-2,117 in existing tools, maintains 98.28% accuracy on obfuscated code, and identified 219 previously unknown malicious packages in real-world deployment. The behavioral patterns show cross-ecosystem applicability with 98.07% accuracy on NPM packages, demonstrating that semantic understanding enables knowledge transfer across programming languages.

cs.CR↗

arXiv2026

HCVR Scene Generation: High Compatibility Virtual Reality Environment Generation for Extended Redirected Walking

Natural walking enhances immersion in virtual environments (VEs), but physical space limitations and obstacles hinder exploration, especially in large virtual scenes. Redirected Walking (RDW) techniques mitigate this by subtly manipulating the virtual camera to guide users away from physical collisions within pre-defined VEs. However, RDW efficacy diminishes significantly when substantial geometric divergence exists between the physical and virtual environments, leading to unavoidable collisions. Existing scene generation methods primarily focus on object relationships or layout aesthetics, often neglecting the crucial aspect of physical compatibility required for effective RDW. To address this, we introduce HCVR (High Compatibility Virtual Reality Environment Generation), a novel framework that generates virtual scenes inherently optimized for alignment-based RDW controllers. HCVR first employs ENI++, a novel, boundary-sensitive metric to evaluate the incompatibility between physical and virtual spaces by comparing rotation-sensitive visibility polygons. Guided by the ENI++ compatibility map and user prompts, HCVR utilizes a Large Language Model (LLM) for context-aware 3D asset retrieval and initial layout generation. The framework then strategically adjusts object selection, scaling, and placement to maximize coverage of virtually incompatible regions, effectively guiding users towards RDW-feasible paths. User studies evaluating physical collisions and layout quality demonstrate HCVR's effectiveness with HCVR-generated scenes, resulting in 22.78 times fewer physical collisions and received 35.89\% less on ENI++ score compared to LLM-based generation with RDW, while also receiving 12.5\% higher scores on user feedback to layout design.

cs.MM↗

arXiv2026

Global Context Compression with Interleaved Vision-Text Transformation

Recent achievements of vision-language models in end-to-end OCR point to a new avenue for low-loss compression of textual information. This motivates earlier works that render the Transformer's input into images for prefilling, which effectively reduces the number of tokens through visual encoding, thereby alleviating the quadratically increased Attention computations. However, this partial compression fails to save computational or memory costs at token-by-token inference. In this paper, we investigate global context compression, which saves tokens at both prefilling and inference stages. Consequently, we propose VIST2, a novel Transformer that interleaves input text chunks alongside their visual encoding, while depending exclusively on visual tokens in the pre-context to predict the next text token distribution. Around this idea, we render text chunks into sketch images and train VIST2 in multiple stages, starting from curriculum-scheduled pretraining for optical language modeling, followed by modal-interleaved instruction tuning. We conduct extensive experiments using VIST2 families scaled from 0.6B to 8B to explore the training recipe and hyperparameters. With a 4$\times$ compression ratio, the resulting models demonstrate significant superiority over baselines on long writing tasks, achieving, on average, a 3$\times$ speedup in first-token generation, 77% reduction in memory usage, and 74% reduction in FLOPS. Our codes and datasets will be public to support further studies.

cs.CV↗

arXiv2025

CogMem: A Cognitive Memory Architecture for Sustained Multi-Turn Reasoning in Large Language Models

Large language models (LLMs) excel at single-turn reasoning but often lose accuracy and coherence over extended, multi-turn interactions. Recent evaluations such as TurnBench highlight recurring failure modes-reasoning bias, task drift, hallucination, overconfidence, and memory decay. Current approaches typically append full conversational histories, causing unbounded context growth, higher computational costs, and degraded reasoning efficiency. We introduce CogMem, a cognitively inspired, memory-augmented LLM architecture that supports sustained iterative reasoning through structured, persistent memory. CogMem incorporates three layers: a Long-Term Memory (LTM) that consolidates cross-session reasoning strategies; a Direct Access (DA) memory that maintains session-level notes and retrieves relevant long-term memories; and a Focus of Attention (FoA) mechanism that dynamically reconstructs concise, task-relevant context at each turn. Experiments on TurnBench show that this layered design mitigates reasoning failures, controls context growth, and improves consistency across extended reasoning chains, moving toward more reliable, human-like reasoning in LLMs.

cs.CL↗

arXiv2025

UCRBench: Benchmarking LLMs on Use Case Recovery

Use cases are widely employed to specify functional requirements, yet existing benchmarks are scarce and face the risk of being misaligned with actual system behavior, similarly limiting the rigorous evaluation of large language models (LLMs) in generating use cases from source code. We address this gap by introducing code-aligned use case benchmarks, constructed through manual validation of both user-goal and subfunction use cases across nine real-world software projects. Using this benchmark, we conduct the first systematic study of LLMs and propose a hierarchical evaluation protocol that assesses actor correctness, name accuracy, path fidelity, and behavioral coverage. The results show that while LLMs can partially reconstruct system functionality, their performance varies significantly across projects, with particularly noticeable shortcomings in domain-specific and multi-module systems. The models also exhibit high omission rates and struggle to maintain consistent abstraction when aggregating subfunctions into user-goal use cases, highlighting both the potential and current limitations of LLM-based use case reverse engineering.

cs.SE↗

arXiv2025

Convergence Dynamics of Over-Parameterized Score Matching for a Single Gaussian

Score matching has become a central training objective in modern generative modeling, particularly in diffusion models, where it is used to learn high-dimensional data distributions through the estimation of score functions. Despite its empirical success, the theoretical understanding of the optimization behavior of score matching, particularly in over-parameterized regimes, remains limited. In this work, we study gradient descent for training over-parameterized models to learn a single Gaussian distribution. Specifically, we use a student model with $n$ learnable parameters and train it on data generated from a single ground-truth Gaussian using the population score matching objective. We analyze the optimization dynamics under multiple regimes. When the noise scale is sufficiently large, we prove a global convergence result for gradient descent. In the low-noise regime, we identify the existence of a stationary point, highlighting the difficulty of proving global convergence in this case. Nevertheless, we show convergence under certain initialization conditions: when the parameters are initialized to be exponentially small, gradient descent ensures convergence of all parameters to the ground truth. We further prove that without the exponentially small initialization, the parameters may not converge to the ground truth. Finally, we consider the case where parameters are randomly initialized from a Gaussian distribution far from the ground truth. We prove that, with high probability, only one parameter converges while the others diverge, yet the loss still converges to zero with a $1/τ$ rate, where $τ$ is the number of iterations. We also establish a nearly matching lower bound on the convergence rate in this regime. This is the first work to establish global convergence guarantees for Gaussian mixtures with at least three components under the score matching framework.

cs.LG↗

arXiv2025

TurnBench-MS: A Benchmark for Evaluating Multi-Turn, Multi-Step Reasoning in Large Language Models

Despite impressive advances in large language models (LLMs), existing benchmarks often focus on single-turn or single-step tasks, failing to capture the kind of iterative reasoning required in real-world settings. To address this limitation, we introduce TurnBench, a novel benchmark that evaluates multi-turn, multi-step reasoning through an interactive code-breaking task inspired by the "Turing Machine Board Game." In each episode, a model must uncover hidden logical or arithmetic rules by making sequential guesses, receiving structured feedback, and integrating clues across multiple rounds. This dynamic setup requires models to reason over time, adapt based on past information, and maintain consistency across steps-capabilities underexplored in current benchmarks. TurnBench includes two modes: Classic, which tests standard reasoning, and Nightmare, which introduces increased complexity and requires robust inferential chains. To support fine-grained analysis, we provide ground-truth annotations for intermediate reasoning steps. Our evaluation of state-of-the-art LLMs reveals significant gaps: the best model achieves 84% accuracy in Classic mode, but performance drops to 18% in Nightmare mode. In contrast, human participants achieve 100% in both, underscoring the challenge TurnBench poses to current models. By incorporating feedback loops and hiding task rules, TurnBench reduces contamination risks and provides a rigorous testbed for diagnosing and advancing multi-step, multi-turn reasoning in LLMs.

cs.CL↗

arXiv2025

From Generation to Detection: A Multimodal Multi-Task Dataset for Benchmarking Health Misinformation

Infodemics and health misinformation have significant negative impact on individuals and society, exacerbating confusion and increasing hesitancy in adopting recommended health measures. Recent advancements in generative AI, capable of producing realistic, human like text and images, have significantly accelerated the spread and expanded the reach of health misinformation, resulting in an alarming surge in its dissemination. To combat the infodemics, most existing work has focused on developing misinformation datasets from social media and fact checking platforms, but has faced limitations in topical coverage, inclusion of AI generation, and accessibility of raw content. To address these issues, we present MM Health, a large scale multimodal misinformation dataset in the health domain consisting of 34,746 news article encompassing both textual and visual information. MM Health includes human-generated multimodal information (5,776 articles) and AI generated multimodal information (28,880 articles) from various SOTA generative AI models. Additionally, We benchmarked our dataset against three tasks (reliability checks, originality checks, and fine-grained AI detection) demonstrating that existing SOTA models struggle to accurately distinguish the reliability and origin of information. Our dataset aims to support the development of misinformation detection across various health scenarios, facilitating the detection of human and machine generated content at multimodal levels.

cs.CL↗