arXiv subjects

Mu Zhou

Publications and source records attributed to Mu Zhou.

At least 19 recordsLinked to original sources

arXiv2026

Act More, Decide Less: Skill-Guided Adaptive Action Chunking for Long-Horizon LLM Agents

Large language model (LLM) agents for long-horizon interactive tasks typically follow a ReAct-style protocol, issuing one primitive action per LLM round. While this enables frequent replanning, it is inefficient for long-horizon tasks where many rounds are spent on routine action sequences. A natural alternative is to let the agent emit variable-length action chunks. However, naively training such policies with standard reinforcement learning fails: the agent either collapses to single-action behavior or over-commits to excessively long sequences. Both failures share a common root cause: the inability to learn chunk boundaries. We propose SPACE, which addresses this challenge by distilling chunk-boundary supervision from trajectory-induced programmatic skills. We induce two-level programmatic skills from successful trajectories, where subskill boundaries serve as direct chunk-boundary supervision. This temporal structure is then distilled into a primitive-chunk policy via hybrid on-/off-policy optimization with chunk-aware credit assignment. Experiments on ALFWorld and ScienceWorld show that SPACE improves success rates by 7.0%-31.3% over the strongest baseline in each setting while reducing average LLM decision rounds by up to 78.9%.

cs.LG↗

arXiv2026

Point-gap topology in amorphous non-Hermitian quantum systems

Recent studies have revealed that not only does the correspondence between spectral winding numbers and skin modes break down in non-Hermitian systems, but the energy spectrum itself is highly sensitive to generic perturbations, system size, and boundary conditions. In amorphous non-Hermitian systems, where the positions of lattice sites are uncertain, the spectral instability becomes even more severe, making it difficult to identify stable topological edge states from the eigenvalue spectrum alone. To overcome this challenge, we introduce a correspondence between stable zero-mode singular states and mid-gap states of the energy spectrum in the thermodynamic limit. Because the singular value spectrum is highly robust against small perturbations and variation in size, topological edge states can be reliably probed via singular values even in finite-sized systems. Based on the singular-value decomposition of the Hamiltonian, we construct a topological invariant in real space to characterize the associated topologically protected edge states. Our approach provides a general strategy for exploring point-gap topology in real space and redefine the non-Hermitian skin effect from a new perspective.

quant-ph↗

arXiv2026

Token-Sparse Medical Multimodal Reasoning via Dual-Stream Reinforcement Learning

Vision-language models (VLMs) combining reinforcement learning (RL) ignite remarkable progress in multimodal reasoning, yet still struggle with medical images, which typically exhibit extremely sparse visual evidence to inform clinical decision-making. We recognize that pruning visual tokens outside the grounding region greatly enhances medical reasoning. However, a united RL framework for active visual token pruning (VTP) and medical multimodal reasoning remains unestablished. Here, we propose a dual-stream RL framework, ViToS, to fulfill token pruning and question answering. ViToS trains one policy model with two task branches, where one focuses on grounding while the other conducts token-sparse reasoning after VTP. Furthermore, we solve the coupled policy learning problem by introducing the cross-feedback sequential optimization, avoiding gradient conflict and facilitating convergence of the shared policy model. Evaluated on seven medical benchmarks, our method reduces visual tokens to 77% of the original sequence length while achieving a 108.27% relative performance on Lingshu-7B and 104.16% relative performance on HuatuoGPT-Vision-7B. Overall, ViToS delivers superior performance and inference speedup, establishing an efficient paradigm for medical multimodal reasoning.

cs.CV↗

arXiv2026

SimAMC: A Fast and Accurate Simulator for Resistive Memory-Based Analog Matrix Computing with Non-Idealities

Analog matrix computing (AMC) circuits leverage resistive memory arrays to perform matrix operations in a massively parallel manner, providing an efficient approach for accelerating data-intensive tasks. However, hardware non-idealities severely impact computational accuracy, making early-stage simulation vital for reliable performance estimation and design optimization. While open-loop circuits for matrix-vector multiplication are well-studied, closed-loop AMC circuits, which solve matrix equations, are computationally more complex and substantially more sensitive to non-idealities, complicating their simulation. In this work, we present SimAMC, a simulator for resistive memory-based closed-loop AMC circuits. SimAMC is capable of modeling matrix inversion and eigenvector solving in the presence of key non-idealities, including device programming error, data conversion error, thermal noise, operational amplifier input offset, and interconnect resistance. For real-valued matrix computing circuits, an alternating iterative algorithm is designed. SimAMC's effectiveness is validated through comparison with SPICE, showing excellent agreement while also demonstrating a speedup of several orders of magnitude.

cs.ET↗

arXiv2026

Agent Skill Evaluation and Evolution: Frameworks and Benchmarks

The growth of agent skills has transformed how agentic systems are built, evaluated, and deployed. As skill libraries continue to scale, rigorous evaluation becomes critical to ensuring their utility, quality, and safety in real-world applications. Consequently, the field is undergoing an emerging paradigm shift from isolated skill creation to automated, evaluation-driven skill evolution. In this survey, we systematically examine the landscape of skill evolution and evaluation beyond foundational skill creation. We categorize evolution into four distinct paradigms, spanning execution feedback, trajectory distillation, compression, and reinforcement learning, showing how each element contributes to improving skill utility and reliability. We also provide an analysis of six skill-centric benchmark categories, identifying structural gaps in benchmark coverage, trade-offs, and metric richness to advance skill research. Finally, we identify open directions for building skill ecosystems that are generalizable, efficient, and verifiably safe. The project URL is https://github.com/Cassie07/AgentSkill_Survey

cs.CL↗

arXiv2026

Alignment Collapse Under KV Cache Quantization: Diagnosis and Mitigation

Key-value (KV) cache quantization is widely used to reduce Large Language Model (LLM) inference memory, yet existing evaluations solely focus on measuring perplexity and accuracy without assessing the safety impact. In this study, we explore alignment preservation under KV cache quantization. Across eleven instruction-tuned models (3.8B-72B) and five benchmarks (1,894 prompts), we find that low-bit quantization can silently destroy safety alignment: Mistral-7B loses 15.2% of its refusals at only 1.03x perplexity, and no universal safe bit-width exists, with sharp model-specific phase transitions invisible to standard metrics. We identify that the root cause is geometric: safety features occupy a low-dimensional activation subspace 10^2-10^3x more vulnerable to quantization noise than the full representation space perplexity averages over. Inspired by this observation, we propose Per-Channel Reduction (PCR), a diagnostic that classifies each model into one of three mechanistic failure modes: outlier-crushes-safety, where safety lives in non-outlier channels collaterally damaged by outlier-driven scale factors; outlier-as-safety, where safety overlaps outlier channels and finer granularity cannot rescue it; and multi-layer dilution, where safety is distributed across many layers and per-layer fixes fail. PCR predicts the correct mitigation direction on all nine primary models and one held-out model from an independent family using 20 calibration prompts. PCR generalizes across unseen prompts, models, and production quantizers, including KIVI with up to 97.2% recovery, succeeding where attention-based allocation methods fail. The resulting training-free protocol, requiring approximately 35 GPU-minutes, recovers up to 97% of lost alignment at minimal memory overhead, addressing vulnerabilities confirmed in production vLLM serving with FP8 KV cache on NVIDIA GPUs.

cs.LG↗

arXiv2026

Evidence Over Plans: Online Trajectory Verification for Skill Distillation

Agent skills can remarkably improve task success rates by using human-written procedural documents, but their quality is difficult to assess without environment-grounded verification. Existing skill generation methods heavily rely on preference logs rather than direct environment interaction, often yielding negligible or even degraded gains. We identify that it is a fundamental timing bottleneck: robust skills should be posterior-based, distilled from empirical environment interaction rather than prior plans. In this study, we introduce the Posterior Distillation Index (PDI), a trajectory-level metric that quantifies how well a distilled skill is grounded in the task-environment evidence. To operationalize PDI, we present SPARK (Structured Pipelines for Autonomous Runnable tasKs and sKill generation) for preserving task execution evidence towards full trajectory-level analysis. SPARK generates environment-verified trajectories used to compute PDI, and it applies PDI as an online diagnostic and intervention signal to ensure posterior skill formation. Across 86 runnable tasks, SPARK-generated skills consistently surpass no-skill baselines and outperform human-written skills on student models (inference cost up to 1,000x cheaper than teacher models). These findings show that PDI-guided distillation produces efficient and transferable skills grounded in the task-environment interaction. We release our code at https://github.com/EtaYang10th/spark-skills .

cs.AI↗

arXiv2026

LUCID-SAE: Learning Unified Vision-Language Sparse Codes for Interpretable Concept Discovery

Sparse autoencoders (SAEs) offer a natural path toward comparable explanations across different representation spaces. However, current SAEs are trained per modality, producing dictionaries whose features are not directly understandable and whose explanations do not transfer across domains. In this study, we introduce LUCID (Learning Unified vision-language sparse Codes for Interpretable concept Discovery), a unified vision-language sparse autoencoder that learns a shared latent dictionary for image patch and text token representations, while reserving private capacity for modality-specific details. We achieve feature alignment by coupling the shared codes with a learned optimal transport matching objective without the need of labeling. LUCID yields interpretable shared features that support patch-level grounding, establish cross-modal neuron correspondence, and enhance robustness against the concept clustering problem in similarity-based evaluation. Leveraging the alignment properties, we develop an automated dictionary interpretation pipeline based on term clustering without manual observations. Our analysis reveals that LUCID's shared features capture diverse semantic categories beyond objects, including actions, attributes, and abstract concepts, demonstrating a comprehensive approach to interpretable multimodal representations.

cs.CV↗

arXiv2026

A Privacy-Preserving Localization Scheme with Node Selection in Mobile Networks

Localization in mobile networks has been widely applied in many scenarios. However, an entity responsible for location estimation exposes both the target and anchors to potential location leakage at any time, creating serious security risks. Although existing studies have proposed privacy-preserving localization algorithms, they still face challenges of insufficient positioning accuracy and excessive communication overhead. In this article, we propose a privacy-preserving localization scheme, named PPLZN. PPLZN protects protects the location privacy of both the target and anchor nodes in crowdsourced localization. Simulation results validate the effectiveness of PPLZN. Evidently, it can achieve accurate position estimation without location leakage and outperform state-of-the-art approaches in both positioning accuracy and communication overhead. In addition, PPLZN significantly reduces computational and communication overhead in large-scale deployments, making it well-fitted for practical privacy-preserving localization in resource-constrained networks.

cs.CR↗

arXiv2025

Modeling Closed-loop Analog Matrix Computing Circuits with Interconnect Resistance

Analog matrix computing (AMC) circuits based on resistive random-access memory (RRAM) have shown strong potential for accelerating matrix operations. However, as matrix size grows, interconnect resistance increasingly degrades computational accuracy and limits circuit scalability. Modeling and evaluating these effects are therefore critical for developing effective mitigation strategies. Traditional SPICE (Simulation Program with Integrated Circuit Emphasis) simulators, which rely on modified nodal analysis, become prohibitively slow for large-scale AMC circuits due to the quadratic growth of nodes and feedback connections. In this work, we model AMC circuits with interconnect resistance for two key operations-matrix inversion (INV) and eigenvector computation (EGV), and propose fast solving algorithms tailored for each case. The algorithms exploit the sparsity of the Jacobian matrix, enabling rapid and accurate solutions. Compared to SPICE, they achieve several orders of magnitude acceleration while maintaining high accuracy. We further extend the approach to open-loop matrix-vector multiplication (MVM) circuits, demonstrating similar efficiency gains. Finally, leveraging these fast solvers, we develop a bias-based compensation strategy that reduces interconnect-induced errors by over 50% for INV and 70% for EGV circuits. It also reveals the scaling behavior of the optimal bias with respect to matrix size and interconnect resistance.

cs.ET↗

arXiv2025

Anatomy-VLM: A Fine-grained Vision-Language Model for Medical Interpretation

Accurate disease interpretation from radiology remains challenging due to imaging heterogeneity. Achieving expert-level diagnostic decisions requires integration of subtle image features with clinical knowledge. Yet major vision-language models (VLMs) treat images as holistic entities and overlook fine-grained image details that are vital for disease diagnosis. Clinicians analyze images by utilizing their prior medical knowledge and identify anatomical structures as important region of interests (ROIs). Inspired from this human-centric workflow, we introduce Anatomy-VLM, a fine-grained, vision-language model that incorporates multi-scale information. First, we design a model encoder to localize key anatomical features from entire medical images. Second, these regions are enriched with structured knowledge for contextually-aware interpretation. Finally, the model encoder aligns multi-scale medical information to generate clinically-interpretable disease prediction. Anatomy-VLM achieves outstanding performance on both in- and out-of-distribution datasets. We also validate the performance of Anatomy-VLM on downstream image segmentation tasks, suggesting that its fine-grained alignment captures anatomical and pathology-related knowledge. Furthermore, the Anatomy-VLM's encoder facilitates zero-shot anatomy-wise interpretation, providing its strong expert-level clinical interpretation capabilities.

cs.CV↗

arXiv2025

Language-Enhanced Generative Modeling for Amyloid PET Synthesis from MRI and Blood Biomarkers

Background: Alzheimer's disease (AD) diagnosis heavily relies on amyloid-beta positron emission tomography (Abeta-PET), which is limited by high cost and limited accessibility. This study explores whether Abeta-PET spatial patterns can be predicted from blood-based biomarkers (BBMs) and MRI scans. Methods: We collected Abeta-PET images, T1-weighted MRI scans, and BBMs from 566 participants. A language-enhanced generative model, driven by a large language model (LLM) and multimodal information fusion, was developed to synthesize PET images. Synthesized images were evaluated for image quality, diagnostic consistency, and clinical applicability within a fully automated diagnostic pipeline. Findings: The synthetic PET images closely resemble real PET scans in both structural details (SSIM = 0.920 +/- 0.003) and regional patterns (Pearson's r = 0.955 +/- 0.007). Diagnostic outcomes using synthetic PET show high agreement with real PET-based diagnoses (accuracy = 0.80). Using synthetic PET, we developed a fully automatic AD diagnostic pipeline integrating PET synthesis and classification. The synthetic PET-based model (AUC = 0.78) outperforms T1-based (AUC = 0.68) and BBM-based (AUC = 0.73) models, while combining synthetic PET and BBMs further improved performance (AUC = 0.79). Ablation analysis supports the advantages of LLM integration and prompt engineering. Interpretation: Our language-enhanced generative model synthesizes realistic PET images, enhancing the utility of MRI and BBMs for Abeta spatial pattern assessment and improving the diagnostic workflow for Alzheimer's disease.

cs.CV↗

arXiv2025

Think Twice to See More: Iterative Visual Reasoning in Medical VLMs

Medical vision-language models (VLMs) excel at image-text understanding but typically rely on a single-pass reasoning that neglects localized visual cues. In clinical practice, however, human experts iteratively scan, focus, and refine the regions of interest before reaching a final diagnosis. To narrow this machine-human perception gap, we introduce ViTAR, a novel VLM framework that emulates the iterative reasoning process of human experts through a cognitive chain of "think-act-rethink-answer". ViTAR treats medical images as interactive objects, enabling models to engage multi-step visual reasoning. To support this approach, we curate a high-quality instruction dataset comprising 1K interactive examples that encode expert-like diagnostic behaviors. In addition, a 16K visual question answering training data has been curated towards fine-grained visual diagnosis. We introduce a two-stage training strategy that begins with supervised fine-tuning to guide cognitive trajectories, followed by the reinforcement learning to optimize decision-making. Extensive evaluations demonstrate that ViTAR outperforms strong state-of-the-art models. Visual attention analysis reveals that from the "think" to "rethink" rounds, ViTAR increasingly anchors visual grounding to clinically critical regions and maintains high attention allocation to visual tokens during reasoning, providing mechanistic insight into its improved performance. These findings demonstrate that embedding expert-style iterative thinking chains into VLMs enhances both performance and trustworthiness of medical AI.

cs.CV↗

arXiv2025

Mediator-Guided Multi-Agent Collaboration among Open-Source Models for Medical Decision-Making

Complex medical decision-making involves cooperative workflows operated by different clinicians. Designing AI multi-agent systems can expedite and augment human-level clinical decision-making. Existing multi-agent researches primarily focus on language-only tasks, yet their extension to multimodal scenarios remains challenging. A blind combination of diverse vision-language models (VLMs) can amplify an erroneous outcome interpretation. VLMs in general are less capable in instruction following and importantly self-reflection, compared to large language models (LLMs) of comparable sizes. This disparity largely constrains VLMs' ability in cooperative workflows. In this study, we propose MedOrch, a mediator-guided multi-agent collaboration framework for medical multimodal decision-making. MedOrch employs an LLM-based mediator agent that enables multiple VLM-based expert agents to exchange and reflect on their outputs towards collaboration. We utilize multiple open-source general-purpose and domain-specific VLMs instead of costly GPT-series models, revealing the strength of heterogeneous models. We show that the collaboration within distinct VLM-based agents can surpass the capabilities of any individual agent. We validate our approach on five medical vision question answering benchmarks, demonstrating superior collaboration performance without model training. Our findings underscore the value of mediator-guided multi-agent collaboration in advancing medical multimodal intelligence.

cs.AI↗

arXiv2025

Floquet composite Dirac semimetals

Dirac semimetals can be classified into types I, II, and III based on the topological charge of their Dirac points. If a three-dimensional (3D) system can be sliced into a family of kz-dependent normal and topological insulators, type I Dirac points separate a 2D normal insulator from a 2D first-order topological insulator, while type II (III) Dirac points separate a 2D normal (first-order) insulator from a 2D second-order topological insulator. To investigate the effects arising from the interplay of distinct Dirac points, one may wonder whether these Dirac points can coexist in a single system. Here, we propose a scheme to induce composite Dirac semimetals by a special Floquet driving that preserves time-reversal and space-inversion symmetries. A general description is established to characterize Dirac semimetals in Floquet systems. The results show that Dirac semimetals hosting coexisting type I, II, and III Dirac points can be induced by delta-function or harmonic driving. Our results provide a promising new avenue for exploring novel Dirac semimetals.

cond-mat.mes-hall↗

arXiv2025

Aligning Large Language Models with Healthcare Stakeholders: A Pathway to Trustworthy AI Integration

The wide exploration of large language models (LLMs) raises the awareness of alignment between healthcare stakeholder preferences and model outputs. This alignment becomes a crucial foundation to empower the healthcare workflow effectively, safely, and responsibly. Yet the varying behaviors of LLMs may not always match with healthcare stakeholders' knowledge, demands, and values. To enable a human-AI alignment, healthcare stakeholders will need to perform essential roles in guiding and enhancing the performance of LLMs. Human professionals must participate in the entire life cycle of adopting LLM in healthcare, including training data curation, model training, and inference. In this review, we discuss the approaches, tools, and applications of alignments between healthcare stakeholders and LLMs. We demonstrate that LLMs can better follow human values by properly enhancing healthcare knowledge integration, task understanding, and human guidance. We provide outlooks on enhancing the alignment between humans and LLMs to build trustworthy real-world healthcare applications.

cs.CY↗

arXiv2025

Show and Segment: Universal Medical Image Segmentation via In-Context Learning

Medical image segmentation remains challenging due to the vast diversity of anatomical structures, imaging modalities, and segmentation tasks. While deep learning has made significant advances, current approaches struggle to generalize as they require task-specific training or fine-tuning on unseen classes. We present Iris, a novel In-context Reference Image guided Segmentation framework that enables flexible adaptation to novel tasks through the use of reference examples without fine-tuning. At its core, Iris features a lightweight context task encoding module that distills task-specific information from reference context image-label pairs. This rich context embedding information is used to guide the segmentation of target objects. By decoupling task encoding from inference, Iris supports diverse strategies from one-shot inference and context example ensemble to object-level context example retrieval and in-context tuning. Through comprehensive evaluation across twelve datasets, we demonstrate that Iris performs strongly compared to task-specific models on in-distribution tasks. On seven held-out datasets, Iris shows superior generalization to out-of-distribution data and unseen classes. Further, Iris's task encoding module can automatically discover anatomical relationships across datasets and modalities, offering insights into medical objects without explicit anatomical supervision.

cs.CV↗

arXiv2025

MedForge: Building Medical Foundation Models Like Open Source Software Development

Foundational models (FMs) have made significant strides in the healthcare domain. Yet the data silo challenge and privacy concern remain in healthcare systems, hindering safe medical data sharing and collaborative model development among institutions. The collection and curation of scalable clinical datasets increasingly become the bottleneck for training strong FMs. In this study, we propose Medical Foundation Models Merging (MedForge), a cooperative framework enabling a community-driven medical foundation model development, meanwhile preventing the information leakage of raw patient data and mitigating synchronization model development issues across clinical institutions. MedForge offers a bottom-up model construction mechanism by flexibly merging task-specific Low-Rank Adaptation (LoRA) modules, which can adapt to downstream tasks while retaining original model parameters. Through an asynchronous LoRA module integration scheme, the resulting composite model can progressively enhance its comprehensive performance on various clinical tasks. MedForge shows strong performance on multiple clinical datasets (e.g., breast cancer, lung cancer, and colon cancer) collected from different institutions. Our major findings highlight the value of collaborative foundation models in advancing multi-center clinical collaboration effectively and cohesively. Our code is publicly available at https://github.com/TanZheling/MedForge.

cs.LG↗