Searcharxiv⌕ Search

arXiv subjects

Yifan Zhang

Publications and source records attributed to Yifan Zhang.

At least 199 records · Page 11Linked to original sources
arXiv2026

GeoEyes: On-Demand Visual Focusing for Evidence-Grounded Understanding of Ultra-High-Resolution Remote Sensing Imagery

The "thinking-with-images" paradigm enables multimodal large language models (MLLMs) to actively explore visual scenes via zoom-in tools. This is essential for ultra-high-resolution (UHR) remote sensing VQA, where task-relevant cues are sparse and tiny. However, we observe a consistent failure mode in existing zoom-enabled MLLMs: Tool Usage Homogenization, where tool calls collapse into task-agnostic patterns, limiting effective evidence acquisition. To address this, we propose GeoEyes, a staged training framework consisting of (1) a cold-start SFT dataset, UHR Chain-of-Zoom (UHR-CoZ), which covers diverse zooming regimes, and (2) an agentic reinforcement learning method, AdaZoom-GRPO, that explicitly rewards evidence gain and answer improvement during zoom interactions. The resulting model learns on-demand zooming with proper stopping behavior and achieves substantial improvements on UHR remote sensing benchmarks, with 54.23% accuracy on XLRS-Bench.

cs.CV↗
arXiv2026

On the Design of KL-Regularized Policy Gradient Algorithms for LLM Reasoning

Policy gradient algorithms have been successfully applied to enhance the reasoning capabilities of large language models (LLMs). KL regularization is ubiquitous, yet the design surface, choice of KL direction (forward vs. reverse), normalization (normalized vs. unnormalized), and estimator ($k_1/k_2/k_3$), is scattered across the literature and often intertwined with off-policy estimation. We ask a focused question: under the off-policy setting, what weighting is required for each KL variant so that the surrogate we optimize yields the exact gradient of the intended KL-regularized objective? We answer this with a compact, unified derivation we call the Regularized Policy Gradient (RPG) view. RPG (i) unifies normalized and unnormalized KL variants and shows that the widely-used $k_3$ penalty is exactly the unnormalized KL; (ii) specifies conditions under which REINFORCE-style losses with stop-gradient are gradient-equivalent to fully differentiable surrogates; (iii) identifies and corrects an off-policy importance-weighting mismatch in GRPO's KL term; and (iv) introduces RPG-Style Clip, a clipped-importance-sampling step within RPG-REINFORCE that enables stable, off-policy policy-gradient training at scale. On mathematical reasoning benchmarks (AIME24, AIME25), RPG-REINFORCE with RPG-Style Clip improves accuracy by up to $+6$ absolute percentage points over DAPO. We extend our experiments to 8K context length, and RPG-REINFORCE with RPG-Style Clip achieves 52% accuracy on AIME25, surpassing the official Qwen3-4B-Instruct model (47%). Notably, RPG is a stable and scalable RL algorithm for LLM reasoning, realized via (a) a KL-correct objective, (b) clipped importance sampling, and (c) an iterative reference-policy update scheme. Project Page: https://github.com/complex-reasoning/RPG.

cs.LG↗
arXiv2026

Stack of correlated insulating states in bilayer graphene kagome superlattice

Graphene-based systems have emerged as a rich platform for exploring emergent quantum phenomena-including superconductivity, magnetism, and correlated insulating behavior-arising from flat electronic bands that enhance many-body interactions. Realizing such flat bands has thus far relied primarily on moiré graphene superlattices or rhombohedral stacking graphene systems, both of which face challenges in reproducibility and tunability. Here, we introduce an artificial Kagome superlattice in bilayer graphene, engineered via nanopatterning of the dielectric substrate to create a precisely defined and electrostatically tunable periodic potential. Magnetotransport measurements reveal the emergence of a stack of correlated insulating states at moderate superlattice potentials, characteristic of strong electron-electron interactions within Kagome-induced flat bands. As temperature increases, these correlated gaps collapse, signaling the thermal suppression of interaction-driven states. Continuum-model calculations confirm the formation of multiple flat minibands and reproduce the observed evolution of band reconstruction. Our results establish dielectric-patterned graphene superlattices as a robust and controllable architecture for realizing flat-band-induced correlated phenomena beyond moiré systems.

cond-mat.mes-hall↗
arXiv2026

Text Before Vision: Staged Knowledge Injection Matters for Agentic RLVR in Ultra-High-Resolution Remote Sensing Understanding

Multimodal reasoning for ultra-high-resolution (UHR) remote sensing (RS) is usually bottlenecked by visual evidence acquisition: the model necessitates localizing tiny task-relevant regions in massive pixel spaces. While Agentic Reinforcement Learning with Verifiable Rewards (RLVR) using zoom-in tools offers a path forward, we find that standard reinforcement learning struggles to navigate these vast visual spaces without structured domain priors. In this paper, we investigate the interplay between post-training paradigms: comparing Cold-start Supervised Fine-Tuning (SFT), RLVR, and Agentic RLVR on the UHR RS benchmark.Our controlled studies yield a counter-intuitive finding: high-quality Earth-science text-only QA is a primary driver of UHR visual reasoning gains. Despite lacking images, domain-specific text injects the concepts, mechanistic explanations, and decision rules necessary to guide visual evidence retrieval.Based on this, we propose a staged knowledge injection recipe: (1) cold-starting with scalable, knowledge-graph-verified Earth-science text QA to instill reasoning structures;and (2) "pre-warming" on the same hard UHR image-text examples during SFT to stabilize and amplify subsequent tool-based RL. This approach achieves a 60.40% Pass@1 on XLRS-Bench, significantly outperforming larger general purpose models (e.g., GPT-5.2, Gemini 3.0 Pro, Intern-S1) and establishing a new state-of-the-art.

cs.AI↗
arXiv2026

Transferable XAI: Relating Understanding Across Domains with Explanation Transfer

Current Explainable AI (XAI) focuses on explaining a single application, but when encountering related applications, users may rely on their prior understanding from previous explanations. This leads to either overgeneralization and AI overreliance, or burdensome independent memorization. Indeed, related decision tasks can share explanatory factors, but with some notable differences; e.g., body mass index (BMI) affects the risks for heart disease and diabetes at the same rate, but chest pain is more indicative of heart disease. Similarly, models using different attributes for the same task still share signals; e.g., temperature and pressure affect air pollution but in opposite directions due to the ideal gas law. Leveraging transfer of learning, we propose Transferable XAI to enable users to transfer understanding across related domains by explaining the relationship between domain explanations using a general affine transformation framework applied to linear factor explanations. The framework supports explanation transfer across various domain types: translation for data subspace (subsuming prior work on Incremental XAI), scaling for decision task, and mapping for attributes. Focusing on task and attributes domain types, in formative and summative user studies, we investigated how well participants could understand AI decisions from one domain to another. Compared to single-domain and domain-independent explanations, Transferable XAI was the most helpful for understanding the second domain, leading to the best decision faithfulness, factor recall, and ability to relate explanations between domains. This framework contributes to improving the reusability of explanations across related AI applications by explaining factor relationships between subspaces, tasks, and attributes.

cs.HC↗
arXiv2026

Game-Theoretic Co-Evolution for LLM-Based Heuristic Discovery

Large language models (LLMs) have enabled rapid progress in automatic heuristic discovery (AHD), yet most existing methods are predominantly limited by static evaluation against fixed instance distributions, leading to potential overfitting and poor generalization under distributional shifts. We propose Algorithm Space Response Oracles (ASRO), a game-theoretic framework that reframes heuristic discovery as a program level co-evolution between solver and instance generator. ASRO models their interaction as a two-player zero-sum game, maintains growing strategy pools on both sides, and iteratively expands them via LLM-based best-response oracles against mixed opponent meta-strategies, thereby replacing static evaluation with an adaptive, self-generated curriculum. Across multiple combinatorial optimization domains, ASRO consistently outperforms static-training AHD baselines built on the same program search mechanisms, achieving substantially improved generalization and robustness on diverse and out-of-distribution instances.

cs.AI↗
arXiv2026

Symbolic dynamics for certain non-invertible $C^{1+β}$ maps

Let $f$ be a non-invertible $C^{1+β}(β>0)$ map with zero Lyapunov exponents and singularities on a closed Riemannian manifold $M$. We consider the symbolic dynamics of $f$. Combining the techniques in recent works of Sarig, Ovadia and Araujo-Lima-Poletti, we construct a countable Markov partition for the invariant set consisting of summable points of the inverse limit space of $(M, f)$ and show that there exists a finite-to-one symbolic extension for $f$ on the corresponding subset of $M$.

math.DS↗
arXiv2026

Modality-Tailored Age of Information for Multimodal Data in Edge Computing Systems

As Internet of Things (IoT) systems scale and device heterogeneity grows, multimodal data have become ubiquitous. Meanwhile, evaluating the freshness of multimodal data is essential, as stale updates would delay task execution, degrade decision accuracy, and undermine safety in latency-sensitive services. However, existing freshness metrics such as Age of Information (AoI) are not suitable for multimodal data, as they do not capture modality-specific characteristics. In this paper, we propose a metric, namely, Modality-Tailored Age of Information (MAoI), to provide a unified and decision-relevant evaluation of freshness for resource management and policy optimization for multimodal data. This metric integrates modality-specific semantic and temporal characteristics, reflecting both age evolution and content importance for multimodal data in multi-access edge computing (MEC) systems. Then, the closed-form expression of the average MAoI is derived, and an MAoI minimization problem is formulated, where sampling intervals and offloading decisions are optimized with practical energy constraints. To effectively solve this problem, a Joint Sampling Offloading Optimization (JSO) algorithm is proposed to jointly optimize the sampling intervals and offloading decisions. It is a block coordinate descent-based algorithm where an optimal sampling-interval subalgorithm is used to update the sampling intervals, and an interference-aware best-response offloading subalgorithm is proposed to update the offloading decisions alternately. Finally, a comprehensive simulation is performed, confirming that the MAoI metric effectively quantifies multimodal freshness compared to traditional AoI, and the JSO algorithm significantly minimizes the average MAoI compared to state-of-the-art algorithms.

cs.NI↗
arXiv2026

StreamShield: A Production-Proven Resiliency Solution for Apache Flink at ByteDance

Distributed Stream Processing Systems (DSPSs) form the backbone of real-time processing and analytics at ByteDance, where Apache Flink powers one of the largest production clusters worldwide. Ensuring resiliency, the ability to withstand and rapidly recover from failures, together with operational stability, which provides consistent and predictable performance under normal conditions, is essential for meeting strict Service Level Objectives (SLOs). However, achieving resiliency and stability in large-scale production environments remains challenging due to the cluster scale, business diversity, and significant operational overhead. In this work, we present StreamShield, a production-proven resiliency solution deployed in ByteDance's Flink clusters. Designed along complementary perspectives of the engine and cluster, StreamShield introduces key techniques to enhance resiliency, covering runtime optimization, fine-grained fault-tolerance, hybrid replication strategy, and high availability under external systems. Furthermore, StreamShield proposes a robust testing and deployment pipeline that ensures reliability and robustness in production releases. Extensive evaluations on a production cluster demonstrate the efficiency and effectiveness of techniques proposed by StreamShield.

cs.DB↗
arXiv2026

IntraSlice: Towards High-Performance Structural Pruning with Block-Intra PCA for LLMs

Large Language Models (LLMs) achieve strong performance across diverse tasks but face deployment challenges due to their massive size. Structured pruning offers acceleration benefits but leads to significant performance degradation. Recent PCA-based pruning methods have alleviated this issue by retaining key activation components, but are only applied between modules in order to fuse the transformation matrix, which introduces extra parameters and severely disrupts activation distributions due to residual connections. To address these issues, we propose IntraSlice, a framework that applies block-wise module-intra PCA compression pruning. By leveraging the structural characteristics of Transformer modules, we design an approximate PCA method whose transformation matrices can be fully fused into the model without additional parameters. We also introduce a PCA-based global pruning ratio estimator that further considers the distribution of compressed activations, building on conventional module importance. We validate our method on Llama2, Llama3, and Phi series across various language benchmarks. Experimental results demonstrate that our approach achieves superior compression performance compared to recent baselines at the same compression ratio or inference speed.

cs.LG↗
arXiv2026

SMCP: Secure Model Context Protocol

Agentic AI systems built around large language models (LLMs) are moving away from closed, single-model frameworks and toward open ecosystems that connect a variety of agents, external tools, and resources. The Model Context Protocol (MCP) has emerged as a standard to unify tool access, allowing agents to discover, invoke, and coordinate with tools more flexibly. However, as MCP becomes more widely adopted, it also brings a new set of security and privacy challenges. These include risks such as unauthorized access, tool poisoning, prompt injection, privilege escalation, and supply chain attacks, any of which can impact different parts of the protocol workflow. While recent research has examined possible attack surfaces and suggested targeted countermeasures, there is still a lack of systematic, protocol-level security improvements for MCP. To address this, we introduce the Secure Model Context Protocol (SMCP), which builds on MCP by adding unified identity management, robust mutual authentication, ongoing security context propagation, fine-grained policy enforcement, and comprehensive audit logging. In this paper, we present the main components of SMCP, explain how it helps reduce security risks, and illustrate its application with practical examples. We hope that this work will contribute to the development of agentic systems that are not only powerful and adaptable, but also secure and dependable.

cs.CR↗
arXiv2026

PACER: Blockwise Pre-verification for Speculative Decoding with Adaptive Length

Speculative decoding (SD) is a powerful technique for accelerating the inference process of large language models (LLMs) without sacrificing accuracy. Typically, SD employs a small draft model to generate a fixed number of draft tokens, which are then verified in parallel by the target model. However, our experiments reveal that the optimal draft length varies significantly across different decoding steps. This variation suggests that using a fixed draft length limits the potential for further improvements in decoding speed. To address this challenge, we propose Pacer, a novel approach that dynamically controls draft length using a lightweight, trainable pre-verification layer. This layer pre-verifies draft tokens blockwise before they are sent to the target model, allowing the draft model to stop token generation if the blockwise pre-verification fails. We implement Pacer on multiple SD model pairs and evaluate its performance across various benchmarks. Our results demonstrate that Pacer achieves up to 2.66x Speedup over autoregressive decoding and consistently outperforms standard speculative decoding. Furthermore, when integrated with Ouroboros, Pacer attains up to 3.09x Speedup.

cs.CL↗
arXiv2026

From Scores to Steps: Diagnosing and Improving LLM Performance in Evidence-Based Medical Calculations

Large language models (LLMs) have demonstrated promising performance on medical benchmarks; however, their ability to perform medical calculations, a crucial aspect of clinical decision-making, remains underexplored and poorly evaluated. Existing benchmarks often assess only the final answer with a wide numerical tolerance, overlooking systematic reasoning failures and potentially causing serious clinical misjudgments. In this work, we revisit medical calculation evaluation with a stronger focus on clinical trustworthiness. First, we clean and restructure the MedCalc-Bench dataset and propose a new step-by-step evaluation pipeline that independently assesses formula selection, entity extraction, and arithmetic computation. Under this granular framework, the accuracy of GPT-4o drops from 62.7% to 43.6%, revealing errors masked by prior evaluations. Second, we introduce an automatic error analysis framework that generates structured attribution for each failure mode. Human evaluation confirms its alignment with expert judgment, enabling scalable and explainable diagnostics. Finally, we propose a modular agentic pipeline, MedRaC, that combines retrieval-augmented generation and Python-based code execution. Without any fine-tuning, MedRaC improves the accuracy of different LLMs from 16.35% up to 53.19%. Our work highlights the limitations of current benchmark practices and proposes a more clinically faithful methodology. By enabling transparent and transferable reasoning evaluation, we move closer to making LLM-based systems trustworthy for real-world medical applications.

cs.CL↗
arXiv2026

Beyond Imprecise Distance Metrics: Trace-Guided Directed Greybox Fuzzing via LLM-Predicted Call Stacks

Directed greybox fuzzing (DGF) aims to efficiently trigger bugs at specific target locations by prioritizing seeds whose execution paths are more likely to reach the targets. However, existing DGF approaches suffer from imprecise potential estimation due to their reliance on static-analysis-based distance metrics. The over-approximation inherent in static analysis causes many seeds with execution paths irrelevant to vulnerability triggering to be mistakenly prioritized, significantly reducing fuzzing efficiency. To address this issue, we propose trace-guided directed greybox fuzzing (TDGF). TDGF replaces static-analysis-based distance metrics with vulnerability-oriented execution information (referred to as guidance traces) to steer directed fuzzing: seeds whose execution paths overlap more with the guidance traces are scheduled earlier for mutation. We empirically study two representative types of guidance traces: the control-flow trace and the call-stack trace of vulnerability-triggering executions. We find that the fine-grained control-flow traces offer nearly the same guidance capability as the coarse-grained call-stack traces, while call-stack traces are also easier for large language models (LLMs) to predict. Based on this insight, we further propose a framework that leverages LLMs to predict the call stack at vulnerability-triggering time and uses it to guide DGF. We implement our approach and evaluate it against several state-of-the-art fuzzers with experiments totaling 58.4 CPU-years. On a suite of real-world programs, our approach triggers vulnerabilities 2.13$\times$ to 3.14$\times$ faster than the baselines. Moreover, through directed patch testing on the latest program versions used in our controlled experiments, our approach discovers 10 new vulnerabilities and 2 incomplete fixes, with 10 assigned CVE IDs.

cs.CR↗
arXiv2026

SADER: Structure-Aware Diffusion Framework with DEterministic Resampling for Multi-Temporal Remote Sensing Cloud Removal

Cloud contamination severely degrades the usability of remote sensing imagery and poses a fundamental challenge for downstream Earth observation tasks. Recently, diffusion-based models have emerged as a dominant paradigm for remote sensing cloud removal due to their strong generative capability and stable optimization. However, existing diffusion-based approaches often suffer from limited sampling efficiency and insufficient exploitation of structural and temporal priors in multi-temporal remote sensing scenarios. In this work, we propose SADER, a structure-aware diffusion framework for multi-temporal remote sensing cloud removal. SADER first develops a scalable Multi-Temporal Conditional Diffusion Network (MTCDN) to fully capture multi-temporal and multimodal correlations via temporal fusion and hybrid attention. Then, a cloud-aware attention loss is introduced to emphasize cloud-dominated regions by accounting for cloud thickness and brightness discrepancies. In addition, a deterministic resampling strategy is designed for continuous diffusion models to iteratively refine samples under fixed sampling steps by replacing outliers through guided correction. Extensive experiments on multiple multi-temporal datasets demonstrate that SADER consistently outperforms state-of-the-art cloud removal methods across all evaluation metrics. The code of SADER is publicly available at https://github.com/zyfzs0/SADER.

cs.CV↗
arXiv2026

SAIR: Cost-Efficient Multi-Stage ML Pipeline Autoscaling via In-Context Reinforcement Learning

Multi-stage ML inference pipelines are difficult to autoscale due to heterogeneous resources, cross-stage coupling, and dynamic bottleneck migration. We present SAIR, an autoscaling framework that uses an LLM as an in-context reinforcement learning controller, improving its policy online from reward-labeled interaction histories without gradient updates. SAIR combines Pareto-dominance reward shaping with a provable separation margin, surprisal-guided experience retrieval for context efficiency, and fine-grained GPU rate control via user-space CUDA interception. We provide regret analysis decomposing error into retrieval coverage and LLM selection components. On four ML serving pipelines under three workload patterns, SAIR achieves the best or tied-best P99 latency and effective resource cost among deployed baselines, improving P99 by up to 50% and reducing effective cost by up to 97% (under GPU rate-control assumptions), with 86% bottleneck detection accuracy and no offline training.

cs.LG↗
arXiv2026

AmbShield: Enhancing Physical Layer Security with Ambient Backscatter Devices against Eavesdroppers

Passive eavesdropping compromises confidentiality in wireless networks, especially in resource-constrained environments where heavyweight cryptography is impractical. Physical layer security (PLS) exploits channel randomness and spatial selectivity to confine information to an intended receiver with modest overhead. However, typical PLS techniques, such as using beamforming, artificial noise, and reconfigurable intelligent surfaces, often involve added active power or specialized deployment, and, in many designs, rely on precise time synchronization and perfect CSI estimation, which limits their practicality. To this end, we propose AmbShield, an AmBD-assisted PLS scheme that leverages naturally distributed AmBDs to simultaneously strengthen the legitimate channel and degrade eavesdroppers' without requiring extra transmit power and with minimal deployment overhead. In AmbShield, AmBDs are exploited as friendly jammers that randomly backscatter to create interference at eavesdroppers, and as passive relays that backscatter the desired signal to enhance the capacity of legitimate devices. We further develop a unified analytical framework that analyzes the exact probability density function (PDF) and cumulative distribution function (CDF) of legitimate and eavesdropper signal-to-interference-noise ratio (SINR), and a closed-form secrecy outage probability (SOP). The analysis provides clear design guidelines on various practical system parameters to minimize SOP. Extensive experiments that include Monte Carlo simulations, theoretical derivations, and high-SNR asymptotic analysis demonstrate the security gains of AmbShield across diverse system parameters under imperfect synchronization and CSI estimation.

cs.CR↗
arXiv2026

StackPilot: Autonomous Function Agents for Scalable and Environment-Free Code Execution

Recent advances in large language models (LLMs) have substantially enhanced automated code generation across a wide range of programming languages. Nonetheless, verifying the correctness and executability of LLM-generated code remains a significant challenge, as traditional methods rely on language-specific compilers and environment-dependent runtimes. To overcome these limitations, we introduce StackPilot, an LLM-native, multi-agent framework designed for language-agnostic code verification and execution, which operates independently of conventional toolchains. StackPilot offers three principal innovations: (1) a Function-as-Agents paradigm, in which each function is modeled as an autonomous agent capable of fine-grained reasoning and collaborative verification; (2) an LLM-as-Executor strategy, which enables scalable verification via stack-based scheduling; and (3) a novel snapshot mechanism that preserves complete execution contexts, facilitating deterministic and lossless context switching during verification. Empirical evaluations demonstrate that StackPilot achieves framework reliability rates between 89% and 97%, substantially outperforming baseline approaches. These results indicate that StackPilot can reliably verify and execute a significantly larger proportion of LLM-generated code across diverse programming tasks compared to existing methods.

cs.PL↗