Beyond Aggregate Risk: Role-Stratified Conformal Risk Control for LLM Tool Calls

Language-model agents act through structured tool calls whose arguments carry different risks. Untrusted content may safely influence an email body but should not determine a recipient, account, command, or credential. Existing statistical methods typically control risk over the entire action, allowing failures in rare, high-risk fields to be obscured by benign arguments. We introduce role-stratified per-field conformal risk control, a calibration layer that wraps any per-field detector and sets separate thresholds and risk budgets for semantic argument roles. For a role with prevalence $p_r$, aggregate-only certification must use an effective budget of $αp_r$ to guarantee role-specific risk $α$, whereas role-stratified calibration certifies each sufficiently sampled role directly with a finite-sample guarantee; rarer roles are handled by pooled certification. Across AgentDojo and InjecAgent with six language models, the empirical utility gap tracks this predicted price of coarseness, and our method achieves the most consistent role-specific budget compliance under model and attack transfer, detector noise, gradual drift, unseen tool suites, and adaptive attacks. It provides formal per-role guarantees under exchangeability or after recalibration, and empirical compliance under frozen distribution shift. These results suggest that structured tool calls should be certified at the semantic-role level, not the whole action.

BAT-RM: A Boundary-Aware Transformer with Region-Aware Multi-Directional Mamba for Clinically Deployed Cervical Cancer Radiotherapy Auto-Contouring

We present a clinically deployed end-to-end auto-contouring system for cervical cancer radiotherapy planning, anchored by the Boundary-Aware Transformer with Region-Aware Mamba (BAT-RM), a hybrid architecture that integrates Sobel-gated boundary attention, a linear-time, multi-directional Mamba module for long-range context, and a boundary-skeleton-guided fusion gate. This design achieves linear-time complexity for long-range context modeling, avoiding the quadratic cost of full spatial self-attention. The full pipeline spans multi-institutional data collection, rigorous inter-rater quality assurance, external validation in an independent cohort, and a web-based clinical interface natively compatible with Varian, RayStation, and Monaco. Against four baselines, BAT-RM achieves superior performance across seven anatomical classes, with statistically significant improvements in target volumes, including GTV and CTV, and in organs at risk such as the rectum and bladder. A prospective multi-center reader study involving 13 radiation oncologists demonstrated that AI assistance elevates junior oncologists' IoU from 0.899 to 0.965, approaching senior-level accuracy, while reducing contouring time by more than 80%. The system also reduced expert consultation rates and improved inter-reader consistency, reflecting gains in both efficiency and quality assurance. Following clinical deployment at a partner hospital, the system reduced patient wait times from days to hours without additional staffing, enabling same-day or next-day initiation of treatment for routine cases. BAT-RM demonstrates that a rigorous research pipeline, from data curation to clinical deployment, can translate directly into measurable patient benefit in resource-constrained settings where the demand for radiotherapy far exceeds specialist capacity.

Equivariant Quantum Clustering with Differential Privacy: Parameter-Efficient Privacy-Preserving Analysis Across Heterogeneous Sensitive Datasets

Privacy-preserving clustering is critical for analyzing sensitive data in healthcare, cybersecurity, and enterprise applications, where maintaining data confidentiality must be balanced with analytical performance. This paper presents Equivariant Quantum Clustering (EQC), a parameter-efficient framework that integrates symmetry-aware quantum circuits with differential privacy to improve the privacy-utility tradeoff. EQC employs p4m equivariant parameter sharing to reduce circuit complexity while preserving informative feature representations. The framework is evaluated on three privacy-sensitive datasets: NSL-KDD, CERT Insider Threat v6.2, and a synthetic MIMIC-III clinical dataset. On the NSL-KDD benchmark, EQC achieves 79.3% clustering accuracy while reducing membership inference attack success to 38.3% under a privacy budget of ε = 1.0 and δ = 10^-5, outperforming representative classical and quantum baselines. Ablation studies indicate that the performance gains primarily arise from parameter-efficient circuit design combined with differential privacy. The results demonstrate that EQC provides a practical quantum-ready framework for secure and privacy-preserving clustering across heterogeneous sensitive datasets.

When Distortion Helps: Secure GNN Precoding with Nonlinear Power Amplifiers

Physical layer security (PLS) provides information-theoretic protection against eavesdropping. While existing techniques assume ideal linear transmitters, power amplifiers (PAs) in practice introduce nonlinear distortion, typically considered detrimental to signal quality. This paper demonstrates that such distortion can instead be exploited as a security asset by redirecting it toward eavesdroppers, particularly in the power-efficient PA saturation regime. To this end, we propose a graph neural network (GNN)-based precoding framework for multi-user multiple-input single-output (MISO) wiretap channels that maximizes the sum secrecy rate by exploiting PA nonlinearity. Since the resulting optimization is highly non-convex, classical methods are intractable. The GNN instead learns precoding strategies directly from legitimate users' channel data, requiring neither eavesdropper channel state information (CSI) nor dedicated artificial noise (AN) power allocation. For this, the Bussgang decomposition and a high-order polynomial PA model provide an analytical secrecy rate as the training objective. At 22 dB signal-to-noise ratio (SNR) under severe PA saturation with input back-off (IBO) $= -1$ dB, the proposed GNN achieves 39.89% and 35.26% higher sum secrecy rate over maximum ratio transmission (MRT) and zero-forcing (ZF), respectively, 17.99% over AN-aided MRT and 8.67% over AN-aided ZF, with 58.13-75.31% lower standard deviation across all baselines.

CANGuard: A Spatio-Temporal CNN-GRU-Attention Hybrid Architecture for Intrusion Detection in In-Vehicle CAN Networks

The Internet of Vehicles (IoV) has become an essential component of smart transportation systems, enabling seamless interaction among vehicles and infrastructure. In recent years, it has played a progressively significant role in enhancing mobility, safety, and transportation efficiency. However, this connectivity introduces severe security vulnerabilities, particularly Denial-of-Service (DoS) and spoofing attacks targeting the Controller Area Network (CAN) bus, which could severely inhibit communication between the critical components of a vehicle, leading to system malfunctions, loss of control, or even endangering passengers' safety. To address this problem, this paper presents CANGuard, a novel spatio-temporal deep learning architecture that combines Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), and an attention mechanism to effectively identify such attacks. The model is trained and evaluated on the CICIoV2024 dataset, achieving competitive performance across accuracy, precision, recall, and F1-score and outperforming existing state-of-the-art methods. A comprehensive ablation study confirms the individual and combined contributions of the CNN, GRU, and attention components. Additionally, a SHAP analysis is conducted to interpret the decision-making process of the model and determine which features have the most significant impact on intrusion detection. The proposed approach demonstrates strong potential for practical and scalable security enhancements in modern IoV environments, thereby ensuring safer and more secure CAN bus communications.

Beyond Dominant Patches: Spatial Credit Redistribution For Grounded Vision-Language Models

Vision-language models (VLMs) frequently hallucinate objects absent from the input image. We trace this failure to spatial credit collapse: activation credit concentrating on sparse visual patches in early transformer layers, which suppresses contextual evidence and increases reliance on language priors. We introduce Spatial Credit Redistribution (SCR), a training-free inference-time intervention that redistributes hidden-state activation from high-attention source patches to their context, guided by low-entropy inputs. We evaluate six model families (Chameleon, LLaVA, and Qwen, including both Qwen-VL and Qwen2-VL) at scales of 7B, 13B, and 30B, on POPE and CHAIR benchmarks. SCR reduces hallucination by ~4.7-6.0 percentage points on POPE-Adversarial, cuts CHAIR-s by 3.7-5.2 percentage points (42-51 percent relative), and CHAIR-i by 2.7-4.4 percentage points (44-58 percent relative), and preserves CIDEr within 0.8 percentage points. Gains are largest for low-entropy inputs, consistent with the theoretical framework. SCR incurs only 43-56 ms overhead (small models: +43-46 ms; large models: +54-56 ms), roughly 3-6 times lower than OPERA and VCD and 1.3-1.7 times lower than OVCD (+72 ms), while Pareto-dominating all three on both hallucination rate and CIDEr, making it practical for real-time settings. A controlled ablation confirms that attention-guided source selection is essential: replacing it with uniform random selection reduces hallucination rate gains from ~4.7-6.0 percentage points to only ~2.6-3.4 percentage points, pointing to credit-collapse as the key driver.

Testbed Evaluation of AI-based Precoding in Distributed MIMO Systems

Distributed MIMO (D-MIMO) has emerged as a key architecture for future sixth-generation (6G) networks, enabling cooperative transmission across spatially distributed access points (APs). However, most existing studies rely on idealized channel models and lack hardware validation, leaving a gap between algorithmic design and practical deployment. Meanwhile, recent advances in artificial intelligence (AI)-driven precoding have shown strong potential for learning nonlinear channel-to-precoder mappings, but their real-world deployment remains limited due to challenges in data collection and model generalization. This work presents a framework for implementing and validating an AI-based precoder on a D-MIMO testbed with hardware reciprocity calibration. A pre-trained graph neural network (GNN)-based model is fine-tuned using real-world channel state information (CSI) collected from the Techtile platform and evaluated under both interpolation and extrapolation scenarios before end-to-end validation. Experimental results demonstrate a 15.7% performance gain over the pre-trained model in the multi-user case after fine-tuning, while in the single-user scenario the model achieves near-maximum ratio transmission (MRT) performance with less than 0.7 bits/channel use degradation out of a total throughput of 5.19 bits/channel use on unseen positions. Further analysis confirms the data efficiency of real-world measurements, showing consistent gains with increasing training samples, and end-to-end validation verifies coherent power focusing comparable to MRT.

GNN-based Precoder Design and Fine-tuning for Cell-free Massive MIMO with Real-world CSI

Cell-free massive MIMO (CF-mMIMO) has emerged as a promising paradigm for delivering uniformly high-quality coverage in future wireless networks. To address the inherent challenges of precoding in such distributed systems, recent studies have explored the use of graph neural network (GNN)-based methods, using their powerful representation capabilities. However, these approaches have predominantly been trained and validated on synthetic datasets, leaving their generalizability to real-world propagation environments largely unverified. In this work, we initially pre-train the GNN using simulated channel state information (CSI) data, which incorporates standard propagation models and small-scale Rayleigh fading. Subsequently, we finetune the model on real-world CSI measurements collected from a physical testbed equipped with distributed access points (APs). To balance the retention of pre-trained features with adaptation to real-world conditions, we adopt a layer-freezing strategy during fine-tuning, wherein several GNN layers are frozen and only the later layers remain trainable. Numerical results demonstrate that the fine-tuned GNN significantly outperforms the pre-trained model, achieving an approximate 8.2 bits per channel use gain at 20 dB signal-to-noise ratio (SNR), corresponding to a 15.7 % improvement. These findings highlight the critical role of transfer learning and underscore the potential of GNN-based precoding techniques to effectively generalize from synthetic to real-world wireless environments.

Cell-free mMIMO Support in the O-RAN Architecture: A PHY Layer Perspective for 5G and Beyond Networks

To keep supporting next-generation requirements, the radio access infrastructure will increasingly densify. Cell-free (CF) network architectures are emerging, combining dense deployments with extreme flexibility in allocating resources to users. In parallel, the Open Radio Access Networks (O-RAN) paradigm is transforming RAN towards an open, intelligent, virtualized, and fully interoperable architecture. This paradigm brings the needed flexibility and intelligent control opportunities for CF networking. In this paper, we document the current O-RAN terminology and contrast it with some common CF processing approaches. We then discuss the main O-RAN innovations and research challenges that remain to be solved.