Papers

We present a perceptually-driven video compression framework integrating implicit neural representations (INRs) and pre-trained video diffusion models to address the extremely low bitrate regime (<0.05 bpp). Our approach exploits the complementary strengths of INRs, which provide a compact video representation, and diffusion models, which offer rich generative priors learned from large-scale datasets. The INR-based conditioning replaces traditional intra-coded keyframes with bit-efficient neural representations trained to estimate latent features and guide the diffusion process. Our joint optimization of INR weights and parameter-efficient adapters for diffusion models allows the model to learn reliable conditioning signals while encoding video-specific information with minimal parameter overhead. Our experiments on UVG, MCL-JCV, and JVET Class-B benchmarks demonstrate substantial improvements in perceptual metrics (LPIPS, DISTS, and FID) at extremely low bitrates, including improvements on BD-LPIPS up to 0.214 and BD-FID up to 91.14 relative to HEVC, while also outperforming VVC and previous strong state-of-the-art neural and INR-only video codecs. Moreover, our analysis shows that INR-conditioned diffusion-based video compression first composes the scene layout and object identities before refining textural accuracy, exposing the semantic-to-visual hierarchy that enables perceptually faithful compression at extremely low bitrates.

This paper investigates the state estimation problem for linear systems subject to Gaussian noise, where the model parameters are unknown. By formulating and solving an optimization problem that incorporates both offline and online system data, a novel data-driven moving horizon estimator (DDMHE) is designed. We prove that the expected 2-norm of the estimation error of the proposed DDMHE is ultimately bounded. Further, we establish an explicit relationship between the system noise covariances and the estimation error of the proposed DDMHE. Moreover, through a sample complexity analysis, we show how the length of the offline data affects the estimation error of the proposed DDMHE. We also quantify the performance gap between the proposed DDMHE using noisy data and the traditional moving horizon estimator with known system matrices. Finally, the theoretical results are validated through numerical simulations.

This paper presents a technique to drive the state of a constrained nonlinear system to a specified target state in finite time, when the system suffers a partial loss in control authority. Our technique builds on a recent method to control constrained nonlinear systems by building a simple, linear driftless approximation at the initial state. We construct a partition of the finite time horizon into successively smaller intervals, and design controlled inputs based on the approximate dynamics in each partition. Under conditions that bound the length of the time horizon, we prove that these inputs result in bounded error from the target state in the original nonlinear system. As successive partitions of the time horizon become shorter, the error reduces to zero despite the effect of uncontrolled inputs. A simulation example on the model of a fighter jet demonstrates that the designed sequence of controlled inputs achieves the target state despite the system suffering a loss of control authority over one of its inputs.

Estimating generation costs from observed electricity market data is essential for market simulation, strategic bidding, and system planning. To that end, we model the relationship between generation costs and production schedules with a latent variable model. Estimating generation costs from observed schedules is then formulated as Bayesian inference. A prior distribution encodes an initial belief on parameters, and the inference consists of updating the belief with the posterior distribution given observations. We use balanced neural posterior estimation (BNPE) to learn this posterior. Validation on the IEEE RTS-96 test system shows that marginal costs are recovered with narrow credible intervals, while start-up costs remain largely unidentifiable from schedules alone. The method is benchmarked against an inverse-optimization algorithm that exhibits larger parameter errors without uncertainty quantification.

Integrated sensing and communication (ISAC) is a key enabler of 6G, supporting environment-aware services. A fundamental sensing task in this setting is reliable multi-target detection and tracking. This paper proposes a temporal graph neural network (TGNN)-based tracking method that exploits delay and Doppler information from the wireless channel. The delay-Doppler map is modeled as a sequence of graphs, and tracking is formulated as a temporal node classification problem, enabling joint clustering and data association of dynamic targets. Using ray-tracing-based channel outputs as ground truth, the method is evaluated across multiple scenes with varying target positions, velocities, and trajectories and is compared with a Kalman filter baseline. Results demonstrate reduced normalized mean squared error (NMSE) in delay and Doppler, leading to more accurate multi-target tracking.

Immunohistochemistry (IHC) is essential for assessing specific immune biomarkers like Human Epidermal growth-factor Receptor 2 (HER2) in breast cancer. However, the traditional protocols of obtaining IHC stains are resource-intensive, time-consuming, and prone to structural damages. Virtual staining has emerged as a scalable alternative, but it faces significant challenges in preserving fine-grained cellular structures while accurately translating biochemical expressions. Current state-of-the-art methods still rely on Generative Adversarial Networks (GANs) or standard convolutional U-Net diffusion models that often struggle with "structure and staining trade-offs". The generated samples are either structurally relevant but blurry, or texturally realistic but have artifacts that compromise their diagnostic use. In this paper, we introduce HistDiT, a novel latent conditional Diffusion Transformer (DiT) architecture that establishes a new benchmark for visual fidelity in virtual histological staining. The novelty introduced in this work is, a) the Dual-Stream Conditioning strategy that explicitly maintains a balance between spatial constraints via VAE-encoded latents and semantic phenotype guidance via UNI embeddings; b) the multi-objective loss function that contributes to sharper images with clear morphological structure; and c) the use of the Structural Correlation Metric (SCM) to focus on the core morphological structure for precise assessment of sample quality. Consequently, our model outperforms existing baselines, as demonstrated through rigorous quantitative and qualitative evaluations.

Model-based multi-agent control requires agents to possess a model of the behavior of others to make strategic decisions. Solution concepts from game theory are often used to model the emergent collective behavior of self-interested agents and have found active use in multi-agent control design. Model predictive games are a class of controllers in which an agent iteratively solves a finite-horizon game to predict the behavior of a multi-agent system and synthesize their own control action. When multiple agents implement these types of controllers, there may exist misspecifications in the respective game models embedded in their controllers, stemming from inaccurate estimates or conjectures of other agents' objectives. This paper analyzes the resulting prediction misalignments and their effects on the system's behavior. We provide criteria for the stability of multi-agent dynamic systems with heterogeneous model predictive game controllers, and quantify the sensitivity of the equilibria to individual agents' game parameters.

Deep image prior (DIP) is an unsupervised deep learning framework that has been successfully applied to a variety of inverse imaging problems. However, DIP-based methods are inherently prone to overfitting, which leads to performance degradation and necessitates early stopping. In this paper, we propose a method to mitigate overfitting in DIP-based hyperspectral image (HSI) denoising by jointly combining robust data fidelity and explicit sensitivity regularization. The proposed approach employs a Smooth $\ell_1$ data term together with a divergence-based regularization and input optimization during training. Experimental results on real HSIs corrupted by Gaussian, sparse, and stripe noise demonstrate that the proposed method effectively prevents overfitting and achieves superior denoising performance compared to state-of-the-art DIP-based HSI denoising methods.

We study the design of an offloaded model predictive control (MPC) operating over a lossy communication channel. We introduce a controller design that utilizes two complementary bandwidth-reduction methods. The first method is a multi-horizon MPC formulation that decreases the number of optimization variables, and therefore the size of transmitted input trajectories. The second method is a communication-rate reduction mechanism that lowers the frequency of packet transmissions. We derive theoretical guarantees on recursive feasibility and constraint satisfaction under minimal assumptions on packet loss, and we establish reference-tracking performance for the rate-reduction strategy. The proposed methods are validated using a hardware-in-the-loop setup with a real 5G network, demonstrating simultaneous improvements in bandwidth efficiency and computational load.

Network slicing is a modern 5G technology that provides efficient network experience for diverse use cases. It is a technique for partitioning a single physical network infrastructure into multiple virtual networks, called slices, each equipped for specific services and requirements. In this work, we particularly deal with radio access network (RAN) slicing and resource allocation to RAN slices. In 5G, physical resource blocks (PRBs) being the fundamental units of radio resources, our main focus is to allocate PRBs to the slices efficiently. While addressing a spectrum of needs for multiple services or the same services with multi-priorities, we need to ensure two vital system properties: i) fairness to every service type (i.e., providing the required resources and a desired range of throughput) even after prioritizing a particular service type, and ii) PRB-optimality or minimizing the unused PRBs in slices. These serve as the core performance evaluation metrics for PRB-allocation in our work. We adopt the 3-layered hierarchical PRB-partitioning technique for allocating PRBs to network slices. The case-specific, AI-based solution of the state-of-the-art method lacks sufficient correctness to ensure consistent system performance. To achieve guaranteed correctness and completeness, we leverage formal methods and propose the first approach for a fair and optimal PRB distribution to RAN slices. We formally model the PRB-allocation problem as a 3-layered framework, FORSLICE, specifically by employing satisfiability modulo theories. Next, we apply formal verification to ensure that the desired system properties: fairness and PRB-optimality, are satisfied by the model. The proposed method offers an efficient, versatile and automated approach compatible with all 3-layered hierarchical network structure configurations, yielding significant system property improvements compared to the baseline.

This paper describes a multi-region control framework for floating offshore wind farms. Specifically, we propose a novel generator torque controller that regulates rotor speed in Region 2, corresponding to wind speeds between the cut-in and rated values. In Region 3 (wind speeds at or above rated but below cut-out speed) we employ a PI-LQR for collective blade pitch. Control blending across the transitional wind speeds (Region 2.5) employs a sigmoid weighting function applied to the control variables. Two modeling paradigms are proposed for farm-level power tracking with rotor speed regularization: a nonlinear model predictive controller (NL-MPC) with a dynamic wake model, and a reduced order model predictive controller based on linear parameter varying turbine models with a time delay representation of wake advection (LPVTD-MPC). These approaches are evaluated over three wind inlet conditions using the PJM ancillary service certification criteria for participation in a secondary frequency regulation market. Results show that both approaches achieve scores of at least 89.9% for the three different testing scenarios, which are well above the qualification threshold of 75%. However, the LPVTD-MPC approach solves the problem in under half the time versus NL-MPC but with slightly larger fluctuations in farm-level power output, highlighting the trade-off between performance and computational tractability. The control framework is among the first to address multi-region wind turbine dynamics together with market driven power tracking objectives for floating offshore wind farms. Such multi-region control becomes increasingly necessary in the floating turbine setting where large (region spanning) wind speed variations are common due to wave induced platform pitching.

The competency of any intelligent agent is bounded by its formal account of the world in which it operates. Clinical AI lacks such an account. Existing frameworks address evaluation, regulation, or system design in isolation, without a shared model of the clinical world to connect them. We introduce the Clinical World Model, a framework that formalizes care as a tripartite interaction among Patient, Provider, and Ecosystem. To formalize how any agent, whether human or artificial, transforms information into clinical action, we develop parallel decision-making architectures for providers, patients, and AI agents, grounded in validated principles of clinical cognition. The Clinical AI Skill-Mix operationalizes competency through eight dimensions. Five define the clinical competency space (condition, phase, care setting, provider role, and task) and three specify how AI engages human reasoning (assigned authority, agent facing, and anchoring layer). The combinatorial product of these dimensions yields a space of billions of distinct competency coordinates. A central structural implication is that validation within one coordinate provides minimal evidence for performance in another, rendering the competency space irreducible. The framework supplies a common grammar through which clinical AI can be specified, evaluated, and bounded across stakeholders. By making this structure explicit, the Clinical World Model reframes the field's central question from whether AI works to in which competency coordinates reliability has been demonstrated, and for whom.

Millimeter-wave (mmWave) communication enables high data rates through large bandwidths and highly directional beamforming, but its sensitivity to blockage and mobility makes reliable beam alignment a central challenge. Limited-probing beam management is a fundamental problem in codebook-based mmWave systems, where only a small subset of beams can be evaluated simultaneously, and the serving decision is restricted to the probed set. Under mobility and noisy feedback, this leads to a sequential and partially observable decision problem in which performance depends critically on the quality of the proposed beam candidates. In this paper, we consider limited-probing beam management and develop a history-conditioned discrete denoising diffusion probabilistic model for beam candidate generation. The proposed method learns from logged probing histories a conditional distribution over promising beam indices, which is then used to construct probing candidates online. Numerical analysis shows that the proposed approach consistently achieves better signal-to-noise ratio, beam-miss probability, and conditional probe regret under tight probing budgets compared with strong learning-based and discriminative baselines. The gains are especially pronounced in low-probing regimes, where accurate candidate generation is most critical.

This work explores the potential of integrating an Intelligent Transmissive Surface (ITS) into an antenna array to improve beamforming performance. We show that integrating a moderate number of passive refractive elements into a small antenna array can significantly improve the Weighted Sum Rate (WSR). We investigate the optimization of the WSR under two distinct operational constraints: a Radiated Power (RP) constraint and a Transmitted Power (TP) constraint. Our analysis reveals that the choice between these constraints significantly impacts the design parameters of the ITS-aided array. By contrasting these approaches, we explore critical design and material parameters, including the array geometry, surface loss, and illumination strategies.

In current MIMO mobile communication systems, phase noise can significantly impair performance. To allow for compensation of these impairments, accurate phase noise modeling is necessary. Numerical modeling of the phase noise process at a phase-locked loop (PLL) output is established in the literature and commonly represented by an Ornstein-Uhlenbeck (OU) process. The corresponding spectrum can be represented by a multi-pole/zero model. This work presents a least squares (LS) method for estimating the PLL parameters such as oscillator constants or PLL bandwidth from a measured phase noise spectrum. The method is applied on the MAX2870 and MAX2871 PLL chips and parameter estimates such as oscillator constants and PLL bandwidths are provided. The resulting parameter set enables both time- and frequency-domain numerical simulations.

Ordinary differential equations (ODE's) are a cornerstone of systems and control theory. Accordingly, they are standard material in undergraduate programs in engineering and there is abundant didactic literature about this topic. Yet, the solution methods and formulas prescribed in this didactic literature are unclear about the assumptions behind their derivation and thus about the limits of their applicability. Specifically, smoothness of the input is rarely discussed, even though it is a critical property to define the character of the solutions and the validity of the methods and formulas prescribed. On the other hand, the relationships with the state space representation (SSR) of linear systems is absent from this same literature and only marginally discussed in more advanced texts. In this paper we detail these gaps left behind in the didactic literature, then we provide a formal delimitation of the boundaries of the standard solutions and methods for linear ODE's. Our analysis relies on some key properties of state space representations, so we establish the formal connections between ODEs and SSR's, defining an equivalence between the two that is absent in the literature and is of conceptual interest by itself.

Near-field (NF) integrated sensing and communication (ISAC) is studied for a uniform circular array (UCA) and monostatic sensing. A continuous-time channel model is derived for the spherical wavefront and orthogonal frequency-division multiplexing. A multi-slot Cramèr-Rao lower bound (CRLB) for joint range-angle-Doppler estimation is derived under a slow-motion approximation, showing how the signal delay and wavefront curvature affect the range bound. A transmit beamformer that minimizes the position CRLB is designed via gradient descent, and a maximum-likelihood estimator is obtained by using the Levenberg-Marquardt algorithm initialized with a grid search. Simulations show that increasing the UCA radius improves angle and motion estimation at all distances, but improves the range bound only at distances far inside the Rayleigh distance, and lowers the reliable communication rate.

Deep stochastic state-space models enable Bayesian filtering in nonlinear, partially observed systems but typically assume a fixed latent structure. When this assumption is violated, parameter adaptation alone may result in persistent belief inconsistency. We introduce Cognitive Flexibility (CF) as a representation-level operator that selects latent structures online via an innovation-based predictive score, while preserving the Bayesian filtering recursion. Structural mismatch is formalized as irreducible predictive inconsistency under fixed structure. The resulting belief–structure recursion is shown to be well posed, to exhibit a structural descent property, and to admit finite switching, with reduction to standard Bayesian filtering under correct specification. Experiments on latent-dynamics mismatch, observation-structure shifts, and well-specified regimes confirm that CF improves predictive accuracy under a mismatch while remaining non-intrusive when the model is correctly specified.

In the last decades, energy-based models (EBMs) have become an important class of probabilistic models in which a component of the likelihood is intractable and therefore cannot be evaluated explicitly. Consequently, parameter estimation in EBMs is challenging for conventional inference methods. In this work, we provide a unified framework that connects noise contrastive estimation (NCE), reverse logistic regression (RLR), multiple importance sampling (MIS), and bridge sampling within the context of EBMs. We further show that these methods are equivalent under specific conditions. This unified perspective clarifies relationships among existing methods and enables the development of new estimators, with the potential to improve statistical and computational efficiency. Furthermore, this study helps elucidate the success of NCE in terms of its flexibility and robustness, while also identifying scenarios in which its performance can be further improved. Hence, rather than being a purely descriptive review, this work offers a unifying perspective and additional methodological contributions. The MATLAB code used in the numerical experiments is also made freely available to support the reproducibility of the results.

Resilience in cyber-physical systems of systems (CPSoS) is often assessed using static indices or point-in-time metrics that do not adequately account for the temporal evolution of risk following a disruption. This paper formalizes resilience as a functional of the risk trajectory by modelling risk as a dynamic state variable. It is analytically shown that key resilience properties are structurally determined by maximum deviation (peak) and effective damping, and that cumulative risk exposure depends on their ratio. A simplified energy-dependent system illustrates the resulting differences in peak magnitude, recovery dynamics, and cumulative impact. The proposed approach links resilience assessment to stability properties of dynamic systems and provides a system-theoretically consistent foundation for the analysis of time-dependent resilience in CPSoS.

This paper proposes a complementary filter on SO(3) for attitude estimation from scalar measurements corresponding to projections of known inertial vectors onto body-frame sensing directions. The observer evolves directly on SO(3) and employs a constant-gain innovation tailored to the scalar-output structure. Under suitable persistence-of-excitation conditions, almost-global asymptotic stability is established when at least three inertial vectors are measured along a common body-frame direction. For configurations involving only two scalar measurements, sufficient conditions for asymptotic convergence are derived together with an explicit characterization of the region of attraction. Numerical simulations illustrate the proposed results.

A key task in embedded vision is visual odometry (VO), which estimates camera motion from visual sensors, and it is a core component in many embedded power-constrained systems, from autonomous robots to augmented and virtual reality wearable devices. The newest class of VO systems combines deep learning models with bio-inspired event-based cameras, which are robust to motion blur and lighting conditions. However, state-of-the-art (SoA) event-based VO algorithms require significant memory and computation. For example, the leading approach DEVO requires 733 MB of memory and 155 billion multiply-accumulate (MAC) operations per frame. We present TinyDEVO, an event-based VO deep learning model designed for resource-constrained microcontroller units (MCUs). We deploy TinyDEVO on an ultra-low-power (ULP) 9-core RISC-V-based MCU, achieving a throughput of approximately 1.2 frames per second with an average power consumption of only 86 mW. Thanks to our neural network architectural optimizations and hyperparameter tuning, TinyDEVO reduces the memory footprint by 11.5x (to 63.8 MB) and the number of operations per frame by 29.7x (to 5.2 billion MACs per frame) compared to DEVO, while maintaining an average trajectory error of 27 cm, i.e., only 19 cm higher than DEVO, on three state-of-the-art datasets. Our work demonstrates, for the first time, the feasibility of an event-based VO pipeline on ultra-low-power devices.

ROI (Region of Interest) video selective encryption based on H.265/HEVC is a technology that protects the sensitive regions of videos by perturbing the syntax elements associated with target areas. However, existing methods typically adopt Tile (with a relatively large size) as the minimum encryption unit, which suffers from problems such as inaccurate encryption regions and low encryption precision. This low-precision encryption makes them difficult to apply in sensitive fields such as medicine, military, and remote sensing. In order to address the aforementioned problem, this paper proposes a fine-grained ROI video selective encryption algorithm based on Coding Units (CUs) and prompt segmentation. First, to achieve a more precise ROI acquisition, we present a novel ROI mapping approach based on prompt segmentation. This approach enables precise mapping of ROIs to small $8\times8$ CU levels, significantly enhancing the precision of encrypted regions. Second, we propose a selective encryption scheme based on multiple syntax elements, which distorts syntax elements within high-precision ROI to effectively safeguard ROI security. Finally, we design a diffusion isolation based on Pulse Code Modulation (PCM) mode and MV restriction, applying PCM mode and MV restriction strategy to the affected CU to address encryption diffusion during prediction. The above three strategies break the inherent mechanism of using Tiles in existing ROI encryption and push the fine-grained level of ROI video encryption to the minimum $8\times8$ CU precision. The experimental results demonstrate that the proposed algorithm can accurately segment ROI regions, effectively perturb pixels within these regions, and eliminate the diffusion artifacts introduced by encryption. The method exhibits great potential for application in medical imaging, military surveillance, and remote areas.

In this paper, we consider data-driven reconstruction of unknown inputs for deterministic, linear time-invariant (LTI) multiple-input multiple-output (MIMO) systems. We propose a novel autoregressive estimator based on a constrained least-squares formulation over Hankel matrices, splitting the problem into an output-consistency constraint and an input-history-matching objective. Our method relies on previously recorded input–output data to represent the system, but does not require knowledge of the true input to initialize the algorithm. We show that the proposed estimator is strictly stable if and only if all the invariant zeros of the trajectory-generating system lie strictly inside the unit circle, which can be verified purely from input and output data. This mirrors existing results from model-based input reconstruction and closes the gap between model-based and data-driven settings. Lastly, we provide numerical examples to demonstrate the theoretical results.

This paper presents a detailed measurement campaign and a comprehensive analysis of 15 GHz ultra-massive multiple-input multiple-output (UM-MIMO) channels tailored for the urban microcell (UMi) environment. Channel sounding is performed over 14.875-15.125 GHz using a time-domain platform comprising a 128-element L-shaped transmit array and a 64-element square receive array. Four representative scenarios are investigated, namely near-field line-of-sight (LoS), near-field foliage-shaded, far-field foliage-shaded, and far-field LoS street canyon scenarios, resulting in 81 distinct transmit-receive links. Based on the measured data, conventional channel characteristics, including path loss, power delay angle profiles, delay spread, and angular spread, are characterized, while UM-MIMO-specific phenomena associated with near-field effects, spatial non-stationarity (SNS), and channel hardening (CHD) are quantitatively analyzed. Channel capacity is further evaluated to reveal the effects of different UMi propagation conditions on system performance. The reported results provide empirical support for the new mid-band spectrum (6-24 GHz, including Frequency Range 3 (FR3)) UM-MIMO channel modeling and offer practical guidance for the design and deployment of future sixth-generation (6G) microcell networks.

Integrating large language models (LLMs) into automatic speech recognition (ASR) has become a dominant paradigm. Although recent LLM-based ASR models have shown promising performance on public benchmarks, it remains challenging to balance recognition quality with latency and overhead, while hallucinations further limit real-world deployment. In this study, we revisit LLM-based ASR from an entropy allocation perspective and introduce three metrics to characterize how training paradigms allocate entropy reduction between the speech encoder and the LLM. To remedy entropy-allocation inefficiencies in prevailing approaches, we propose a principled multi-stage training strategy grounded in capability-boundary awareness, optimizing parameter efficiency and hallucination robustness. Specifically, we redesign the pretraining strategy to alleviate the speech-text modality gap, and further introduce an iterative asynchronous SFT stage between alignment and joint SFT to preserve functional decoupling and constrain encoder representation drift. Experiments on Mandarin and English benchmarks show that our method achieves competitive performance with state-of-the-art models using only 2.3B parameters, while also effectively mitigating hallucinations through our decoupling-oriented design.

Multi-Agent Path Finding (MAPF) is a fundamental coordination problem in large-scale robotic and cyber-physical systems, where multiple agents must compute conflict-free trajectories with limited computational and communication resources. While centralised optimal solvers provide guarantees on solution optimality, their exponential computational complexity limits scalability to large-scale systems and real-time applicability. Existing decentralised heuristics are faster, but result in suboptimal outcomes and high cost disparities. This paper proposes a decentralised coordination framework for cooperative MAPF based on Karma mechanisms - artificial, non-tradeable credits that account for agents' past cooperative behaviour and regulate future conflict resolution decisions. The approach formulates conflict resolution as a bilateral negotiation process that enables agents to resolve conflicts through pairwise replanning while promoting long-term fairness under limited communication and without global priority structures. The mechanism is evaluated in a lifelong robotic warehouse multi-agent pickup-and-delivery scenario with kinematic orientation constraints. The results highlight that the Karma mechanism balances replanning effort across agents, reducing disparity in service times without sacrificing overall efficiency. Code: https://github.com/DerKevinRiehl/karma_dmapf

Large language models (LLMs) have recently demonstrated strong potential for autonomous vehicle motion planning by reformulating trajectory prediction as a language generation problem. However, deploying capable LLMs in resource-constrained onboard systems remains a fundamental challenge. In this paper, we study how to effectively transfer motion planning knowledge from a large teacher LLM to a smaller, more deployable student model. We build on the GPT-Driver framework, which represents driving scenes as language prompts and generates waypoint trajectories with chain-of-thought reasoning, and investigate two student training paradigms: (i) on-policy generalized knowledge distillation (GKD), which trains the student on its own self-generated outputs using dense token-level feedback from the teacher, and (ii) a dense-feedback reinforcement learning (RL) baseline that uses the teacher's log-probabilities as per-token reward signals in a policy gradient framework. Experiments on the nuScenes benchmark show that GKD substantially outperforms the RL baseline and closely approaches teacher-level performance despite a 5$\times$ reduction in model size. These results highlight the practical value of on-policy distillation as a principled and effective approach to deploying LLM-based planners in autonomous driving systems.

We propose a Physics Informed Learning framework for reconstructing traffic density from sparse trajectory data. The approach combines a second-order Aw-Rascle and Zhang model with a first-order training stage to estimate the equilibrium velocity. The method is evaluated in both equilibrium and transient traffic regimes using SUMO simulations. Results show that while learning the equilibrium velocity improves reconstruction under steady state conditions, it becomes unstable in transient regimes due to the breakdown of the equilibrium assumption. In contrast, the second-order model consistently provides more accurate and robust reconstructions than first-order approaches, particularly in nonequilibrium conditions.

In the first part of this two-part paper a game-theoretic decentralized real-time control is proposed in the context of Electric Vehicle (EV) Charging Station (CS). This method, relying on a Stackelberg Game-based Alternating Direction of Multipliers (SG-ADMM), intends to steer the EVs' individual objectives towards the CS optimum by means of an incentive design mechanism, while controlling the EV power dispatch in a distributed manner. We integrate SG-ADMM in a hierachical multi-layered Energy Management System (EMS) as the real-time control algorithm, formulating the two-layer approach so that the SG leader (i.e., the CS), holding commitment power, trades off the available power with the incentives to the EVs, and the SG followers (i.e., the EVs) optimizes their charging curve in response to the leader decision. In this second part, we demonstrate the applicability of SG-ADMM as a incentive design mechanism inside an EVCS EMS, testing it in a large-scale EVCS. We benchmark this method with a decentralized (ADMM-based), a centralized and a uncontrolled approach, showing that our method exploits EV-level flexibility in a cost-effective, fair and computationally efficient manner.