Papers

Autonomous vehicles (AVs) rely on sensors and deep neural networks (DNNs) to perceive their surrounding environment and make maneuver decisions in real time. However, achieving real-time DNN inference in the AV's perception pipeline is challenging due to the large gap between the computation requirement and the AV's limited resources. Most, if not all, of existing studies focus on optimizing the DNN inference time to achieve faster perception by compressing the DNN model with pruning and quantization. In contrast, we present a Predictable Perception system with DNNs (PP-DNN) that reduce the amount of image data to be processed while maintaining the same level of accuracy for multi-tenant DNNs by dynamically selecting critical frames and regions of interest (ROIs). PP-DNN is based on our key insight that critical frames and ROIs for AVs vary with the AV's surrounding environment. However, it is challenging to identify and use critical frames and ROIs in multi-tenant DNNs for predictable inference. Given image-frame streams, PP-DNN leverages an ROI generator to identify critical frames and ROIs based on the similarities of consecutive frames and traffic scenarios. PP-DNN then leverages a FLOPs predictor to predict multiply-accumulate operations (MACs) from the dynamic critical frames and ROIs. The ROI scheduler coordinates the processing of critical frames and ROIs with multiple DNN models. Finally, we design a detection predictor for the perception of non-critical frames. We have implemented PP-DNN in an ROS-based AV pipeline and evaluated it with the BDD100K and the nuScenes dataset. PP-DNN is observed to significantly enhance perception predictability, increasing the number of fusion frames by up to 7.3x, reducing the fusion delay by >2.6x and fusion-delay variations by >2.3x, improving detection completeness by 75.4% and the cost-effectiveness by up to 98% over the baseline.

Near-field multi-antenna wireless communication has attracted growing research interest in recent years. Despite this development, most of the current literature on antennas and reflecting structures relies on simplified models, whose validity for real systems remains unclear. In this paper, we introduce a physically consistent near-field model, which we use to evaluate commonly used models. Our results indicate that common models are sufficient for basic beamfocusing, but fail to accurately predict the sidelobes and frequency dependence of reflecting structures.

In this paper, we derive the continuous space-time equations of motion of a three-dimensional geometrically exact rod, or the Cosserat rod, incorporating planar cross-sectional deformation. We then adopt the Lie group variational integrator technique to obtain a discrete model of the rod incorporating both rotational motion and cross-sectional deformation as well. The resulting discrete model possesses several desirable features: it ensures volume conservation of the discrete elements by considering cross-sectional deformation through a local dilatation factor, it demonstrates the beneficial properties associated with the variational integrator technique, such as the preservation of the rotational configuration, and energy conservation with a bounded error. An exhaustive set of numerical results under various initial conditions of the rod demonstrates the efficacy of the model in replicating the physics of the system.

This paper studies a fluid-antenna-enabled integrated bistatic sensing and backscatter communication system for future networks where connectivity, power delivery, and environmental awareness are jointly supported by the same infrastructure. A multi-antenna base station (BS) with transmitting fluid antennas serves downlink users, energizes passive tags, and illuminates radar targets, while a spatially separated multi-antenna reader decodes tag backscatter and processes radar echoes to avoid the strong self-interference that would otherwise obscure weak returns at the BS. The coexistence of tags and targets, however, induces severe near–far disparities and multi-signal interference, which can be mitigated by fluid antennas through additional spatial degrees of freedom that reshape the multi-hop channels. We formulate a transmit-power minimization problem that jointly optimizes the BS information beamformers, sensing covariance matrix, reader receive beamformers, tag reflection coefficients, and fluid-antenna (FA) positions under heterogeneous quality of service constraints for communication, backscatter, and sensing, as well as energy-harvesting and FA geometry requirements. To tackle the resulting non-convex problem, we develop an alternating-optimization block-coordinate framework that solves four tractable subproblems using semidefinite relaxation, majorization–minimization, and successive convex approximation. Numerical results show consistent transmit-power savings over fixed-position antennas and zero-forcing baselines, achieving about 13.7% and 54.5% reductions, respectively.

We define trajectory predictive control (TPC) as a class of indirect data-driven predictive control (DDPC) methods that represent future outputs as linear in past inputs/outputs and future inputs. TPC unifies many DDPC variants with different predictor structures. We introduce a predictor with a state-space representation and show that with it, TPC inherits the mature theory of linear model predictive control. In numerical experiments, the state-space predictor outperforms existing predictors, especially for small training datasets.

Discrete audio tokenizers are fundamental to empowering large language models with native audio processing and generation capabilities. Despite recent progress, existing approaches often rely on pretrained encoders, semantic distillation, or heterogeneous CNN-based architectures. These designs introduce fixed inductive biases that limit reconstruction fidelity and hinder effective scaling. In this paper, we argue that discrete audio tokenization should be learned fully end-to-end using a homogeneous and scalable architecture. To this end, we first propose CAT (Causal Audio Tokenizer with Transformer), a purely Transformer-based architecture that jointly optimizes the encoder, quantizer, and decoder from scratch for high-fidelity reconstruction. Building on the CAT architecture, we develop MOSS-Audio-Tokenizer, a large-scale audio tokenizer featuring 1.6 billion parameters, pre-trained on 3 million hours of diverse, general audio data. We show that this simple, fully end-to-end approach built from homogeneous, causal Transformer blocks scales gracefully and supports high-fidelity reconstruction across diverse audio domains. Across speech, sound, and music, MOSS-Audio-Tokenizer consistently outperforms prior codecs over a wide range of bitrates, while exhibiting predictable improvements with increased scale. Notably, leveraging the discrete tokens from our model, we develop the first purely autoregressive TTS model that surpasses prior non-autoregressive and cascaded systems. Furthermore, MOSS-Audio-Tokenizer enables competitive ASR performance without auxiliary encoders. Our findings position the CAT architecture as a unified, scalable interface for the next generation of native audio foundation models.

Training recurrent neural networks (RNNs) with standard backpropagation through time (BPTT) can be challenging, especially in the presence of long input sequences. A practical alternative to reduce computational and memory overhead is to perform BPTT repeatedly over shorter segments of the training data set, corresponding to truncated BPTT. In this paper, we examine the training of RNNs when using such a truncated learning approach for time series tasks. Specifically, we establish theoretical bounds on the accuracy and performance loss when optimizing over subsequences instead of the full data sequence. This reveals that the burn-in phase of the RNN is an important tuning knob in its training, with significant impact on the performance guarantees. We validate our theoretical results through experiments on standard benchmarks from the fields of system identification and time series forecasting. In all experiments, we observe a strong influence of the burn-in phase on the training process, and proper tuning can lead to a reduction of the prediction error on the training and test data of more than 60% in some cases.

Retinal implants aim to restore functional vision despite photoreceptor degeneration, yet are fundamentally constrained by low resolution electrode arrays and patient-specific perceptual distortions. Most deployed encoders rely on task-agnostic downsampling and linear brightness-to-amplitude mappings, which are suboptimal under realistic perceptual models. While global inverse problems have been formulated as neural networks, such approaches can be fast at inference, and can achieve high reconstruction fidelity, but require training and have limited generalizability to arbitrary inputs. We cast stimulus encoding as a constrained sparse least-squares problem under a linearized perceptual forward model. Our key observation is that the resulting perception matrix can be highly sparse, depending on patient and implant configuration. Building on this, we apply an efficient projected residual norm steepest descent solver that exploits sparsity and supports stimulus bounds via projection. In silico experiments across four simulated patients and implant resolutions from $15\times15$ to $100\times100$ electrodes demonstrate improved reconstruction fidelity, with up to $+0.265$ SSIM increase, $+12.4\,\mathrm{dB}$ PSNR, and $81.4\%$ MAE reduction on Fashion-MNIST compared to Lanczos downsampling.

We apply several machine learning algorithms to the problem of anomaly detection in operational data for large-scale, high-voltage electric power grids. We observe important differences in the performance of the algorithms. Neural networks typically outperform classical algorithms such as k-nearest neighbors and support vector machines, which we explain by the strong contextual nature of the anomalies. We show that unsupervised learning algorithm work remarkably well and that their predictions are robust against simultaneous, concurring anomalies.

Consider a subfield of the field of rational functions in several indeterminates. We present an algorithm that, given a set of generators of such a subfield, finds a simple generating set. We provide an implementation of the algorithm and show that it improves upon the state of the art both in efficiency and the quality of the results. Furthermore, we demonstrate the utility of simplified generators through several case studies from different application domains, such as structural parameter identifiability. The main algorithmic novelties include performing only partial Gröbner basis computation via sparse interpolation and efficient search for polynomials of a fixed degree in a subfield of the rational function field.

Despite the extensive body of work on backstepping for one-dimensional PDEs, results in higher dimensions remain comparatively limited. Most available methods either exploit particular symmetries of the PDE or address problems posed on parallelepiped domains. To the best of our knowledge, the only approach that enables the design of backstepping controllers on non-parallelepiped regions without symmetry assumptions is the domain extension technique. This method, however, presents several drawbacks. In particular, the control input at each time instant is obtained by simulating a PDE on an extended domain, from which the actual input on the original domain is approximated. By contrast, in the one-dimensional setting, once the time-independent backstepping gain kernel is known, the control input can be computed in closed form as a feedback depending solely on the state at that same instant. Moreover, problems such as output-feedback design or adaptive and robust control do not appear straightforward to address with the domain extension method, at least to the best of our knowledge. These considerations motivate the search, whenever possible, for alternatives that preserve the main advantages of one-dimensional backstepping. A motivating example for the domain extension method is the control of the heat equation on a piano-shaped domain, with actuation applied at the tail of the piano. In this extended abstract, we show through a simple calculation that the domain extension method is not required in this setting. Instead, a strategy akin to that used for parallelepiped domains can be adopted. This result constitutes a first instance of a broader framework for backstepping control of asymmetric PDEs posed on non-parallelepiped regions, which we refer to as domains with graph-monotone boundaries. The general framework is developed in a forthcoming paper.

In this paper, we investigate a class of port-Hamiltonian systems with singular vector fields. We show that, under suitable conditions, their interconnection with passive systems ensures convergence to a prescribed non-equilibrium steady state. At first glance, this behavior appears to contradict the seemingly passive structure of port-Hamiltonian systems, since sustaining a non-equilibrium steady state requires continuous power injection. We resolve this apparent paradox by showing that the singularity in the vector field induces a sliding mode that contributes effective energy, enabling maintenance of the steady state and demonstrating that the system is not passive. Furthermore, we consider regularizations of the singular dynamics and show that the resulting systems are cyclo-passive, while still capable of supplying the required steady-state power. These results clarify the role of singularities in port-Hamiltonian systems and provide new insight into their energetic properties.

This study presents an efficient field-programmable gate array (FPGA) implementation of a polynomial spline function-based statistical compression algorithm designed to address the critical challenge of massive data transfer bandwidth in emerging high-spatial-resolution single-photon avalanche diode (SPAD) arrays, where data rates can reach tens of gigabytes per second. In our experiments, the proposed hardware implementation achieves a compression ratio of 512x compared with conventional histogram-based outputs, with the potential for further improvement. The algorithm is first optimized in software using fixed-point (FXP) arithmetic and look-up tables (LUTs) to eliminate explicit additions, multiplications, and non-linear operations. This enables a careful balance between accuracy and hardware resource utilization. Guided by this trade-off analysis, online sketch processing elements (SPEs) are implemented on an FPGA to directly process time-stamp streams from the SPAD sensor. The implementation is validated using a customized LiDAR setup with a 192 x 128-pixel SPAD array. This work demonstrates histogram-free online depth reconstruction with high fidelity, effectively alleviating the time-stamp transfer bottleneck of SPAD arrays and offering scalability as pixel counts continue to increase for future SPADs.

In this paper, the problem of tracking a given reference output trajectory is investigated for the class of Boolean control networks, by resorting to their algebraic representation. First, the case of a finite-length reference trajectory is addressed, and the analysis and algorithm first proposed in [17] are extended to be able to deal with arbitrary initial conditions and to identify all possible solutions. The approach developed for the finite-length case is then adjusted to cope with periodic reference output trajectories. The results of the paper are illustrated through an example.

Deep learning models trained in a supervised setting have revolutionized audio and speech processing. However, their performance inherently depends on the quantity of human-annotated data, making them costly to scale and prone to poor generalization under unseen conditions. To address these challenges, Self-Supervised Learning (SSL) has emerged as a promising paradigm, leveraging vast amounts of unlabeled data to learn relevant representations. The application of SSL for Automatic Speech Recognition (ASR) has been extensively studied, but research on other downstream tasks, notably Speaker Recognition (SR), remains in its early stages. This work describes major SSL instance-invariance frameworks (e.g., SimCLR, MoCo, and DINO), initially developed for computer vision, along with their adaptation to SR. Various SSL methods for SR, proposed in the literature and built upon these frameworks, are also presented. An extensive review of these approaches is then conducted: (1) the effect of the main hyperparameters of SSL frameworks is investigated; (2) the role of SSL components is studied (e.g., data-augmentation, projector, positive sampling); and (3) SSL frameworks are evaluated on SR with in-domain and out-of-domain data, using a consistent experimental setup, and a comprehensive comparison of SSL methods from the literature is provided. Specifically, DINO achieves the best downstream performance and effectively models intra-speaker variability, although it is highly sensitive to hyperparameters and training conditions, while SimCLR and MoCo provide robust alternatives that effectively capture inter-speaker variability and are less prone to collapse. This work aims to highlight recent trends and advancements, identifying current challenges in the field.

The coexistence of terrestrial networks (TN) and non-terrestrial networks (NTN) in the frequency range 3 (FR3) upper mid-band presents considerable interference concerns, as dense TN deployments can severely degrade NTN downlink performance. Existing studies rely on interference-nulling beamforming, precoding, or exclusion zones that require accurate channel state information (CSI) and static coordination, making them unsuitable for dynamic NTN scenarios. To overcome these limitations, we develop an optimization framework that jointly controls TN downlink power, uplink power, and antenna downtilt to protect NTN links while preserving terrestrial performance. The resultant non-convex coupling between TN and NTN parameters is addressed by a Proximal Policy Optimization (PPO)-based reinforcement learning method that develops adaptive power and tilt control strategies. Simulation results demonstrate a reduction up to 8 dB in the median interference-to-noise ratio (INR) while maintaining over 87% TN basestation activity, outperforming conventional baseline methods and validating the feasibility of the proposed strategy for FR3 coexistence.

In signal processing, the data collected from sensing devices is often a noisy linear superposition of multiple components, and the estimation of components of interest constitutes a crucial pre-processing step. In this work, we develop a Bayesian framework for signal component decomposition, which combines Gibbs sampling with plug-and-play (PnP) diffusion priors to draw component samples from the posterior distribution. Unlike many existing methods, our framework supports incorporating component-wise model-driven and data-driven priors into diffusion models in a unified manner. Moreover, the proposed posterior sampler allows component priors to be learned separately and flexibly combined for different decomposition tasks at inference time. Under suitable assumptions, the proposed Diffusion-within-Gibbs (DiG) sampler provably produces samples from the posterior distribution. We also show that DiG can be interpreted as an extension of a class of recently proposed diffusion-based samplers, and that, for suitable classes of sensing operators, DiG better exploits the structure of the measurement model. Numerical experiments demonstrate the superior performance of our method over existing approaches.

Neural network based receivers have recently demonstrated superior system-level performance compared to traditional receivers. However, their practicality is limited by high memory and power requirements, as separate weight sets must be stored for each code rate. To address this challenge, we propose LOREN, a Low Rank-Based Code-Rate Adaptation Neural Receiver that achieves adaptability with minimal overhead. LOREN integrates lightweight low rank adaptation adapters (LOREN adapters) into convolutional layers, freezing a shared base network while training only small adapters per code rate. An end-to-end training framework over 3GPP CDL channels ensures robustness across realistic wireless environments. LOREN achieves comparable or superior performance relative to fully retrained base neural receivers. The hardware implementation of LOREN in 22nm technology shows more than 65% savings in silicon area and up to 15% power reduction when supporting three code rates.

This paper investigates symbol detection for single-carrier communication systems operating in the presence of additive interference with Nakagami-m statistics. Such interference departs from the assumptions underlying conventional detection methods based on Gaussian noise models and leads to detection mismatch that fundamentally affects symbol-level performance. In particular, the presence of random interference amplitude and non-uniform phase alters the structure of the optimal decision regions and renders standard Euclidean distance-based detectors suboptimal. To address this challenge, we develop the maximum-likelihood Gaussian-phase approximate (ML-G) detector, a low-complexity detection rule that accurately approximates maximum-likelihood detection while remaining suitable for practical implementation. The proposed detector explicitly incorporates the statistical properties of the interference and induces decision regions that differ significantly from those arising under conventional metrics. Building on the ML-G framework, we further investigate constellation design under interference-aware detection and formulate an optimization problem that seeks symbol placements that minimize the symbol error probability subject to an average energy constraint. The resulting constellations are obtained numerically and adapt to the interference environment, exhibiting non-standard and asymmetric structures as interference strength increases. Simulation results demonstrate clear symbol error probability gains over established benchmark schemes across a range of interference conditions, particularly in scenarios with dominant additive interference.

This paper explores sparsification methods as a form of regularization in Graph Neural Networks (GNNs) to address high memory usage and computational costs in large-scale graph applications. Using techniques from Network Science and Machine Learning, including Erdős-Rényi for model sparsification, we enhance the efficiency of GNNs for real-world applications. We demonstrate our approach on N-1 contingency assessment in electrical grids, a critical task for ensuring grid reliability. We apply our methods to three datasets of varying sizes, exploring Graph Convolutional Networks (GCN) and Graph Isomorphism Networks (GIN) with different degrees of sparsification and rewiring. Comparison across sparsification levels shows the potential of combining insights from both research fields to improve GNN performance and scalability. Our experiments highlight the importance of tuning sparsity parameters: while sparsity can improve generalization, excessive sparsity may hinder learning of complex patterns. Our adaptive rewiring approach, particularly when combined with early stopping, proves promising by allowing the model to adapt its connectivity structure during training. This research contributes to understanding how sparsity can be effectively leveraged in GNNs for critical applications like power grid reliability analysis.

Cooperative Adaptive Cruise Control (CACC) enables vehicle platooning through inter-vehicle communication, improving traffic efficiency and safety. Conventional CACC relies on feedback linearization, assuming exact knowledge of vehicle parameters. However, the nonlinear longitudinal vehicle dynamics are subject to parametric uncertainty. Applying such feedback linearization with a nominal model yields imperfect cancellation, leading to model mismatch and degraded performance with off-the-shelf CACC controllers. To improve robustness without redesigning the CACC law, we explicitly model the mismatch between the ideal closed-loop dynamics, assumed by the CACC design, and the actual dynamics under parametric uncertainties. Robustness is formulated as an $\mathcal{L}_2$ trajectory-matching problem, minimizing the energy of this mismatch to make the uncertain system behave as closely as possible to the ideal model. This objective is addressed by optimizing over plant equivalent controller (PEC) realizations that preserve the nominal closed-loop behavior while mitigating the effects of parametric uncertainty. Stability and performance are enforced via LMIs, yielding a convex optimization problem applicable to heterogeneous platoons. Experimental results demonstrate improved robustness and performance under parametric uncertainty while preserving nominal behavior.

Active reconfigurable intelligent surfaces (RISs) can mitigate the double-fading loss of passive reflection in satellite downlinks. However, their gains are limited by random co-channel interference, gain-dependent amplifier noise, and regulatory emission constraints. This paper develops a stochastic reliability framework for active RIS-assisted satellite downlinks by modeling the desired and interfering channels, receiver noise, and RIS amplifier noise as random variables. The resulting instantaneous signal-to-interference-plus-noise ratio (SINR) model explicitly captures folded cascaded amplifier noise and reveals a finite high-gain SINR ceiling. To guarantee a target outage level, we formulate a chance-constrained max-SINR design that jointly optimizes the binary RIS configuration and a common amplification gain. The chance constraint is handled using a sample-average approximation (SAA) with a violation budget. The resulting feasibility problem is solved as a mixed-integer second-order cone program (MISOCP) within a bisection search over the SINR threshold. Practical implementation is enforced by restricting the gain to an admissible range determined by small-signal stability and effective isotropic radiated power (EIRP) limits. We also derive realization-wise SINR envelopes based on eigenvalue and l1-norm bounds, which provide interpretable performance limits and fast diagnostics. Monte Carlo results show that these envelopes tightly bound the simulated SINR, reproduce the predicted saturation behavior, and quantify performance degradation as interference increases. Overall, the paper provides a solver-ready, reliability-targeting design methodology whose achieved reliability is validated through out-of-sample Monte Carlo testing under realistic randomness and hardware constraints.

The expansion of the low-altitude economy is contingent on reliable cellular connectivity for unmanned aerial vehicles (UAVs). A key challenge in pre-flight planning is predicting communication link quality along proposed and pre-defined routes, a task hampered by sparse measurements that render existing radio map methods ineffective. This paper introduces a transfer learning framework for high-fidelity route-level radio map prediction. Our key insight is to leverage abundant crowdsourced ground signals as auxiliary supervision. To bridge the significant domain gap between ground and aerial data and address spatial sparsity, our framework learns general propagation priors from simulation, performs adversarial alignment of the feature spaces, and is fine-tuned on limited real UAV measurements. Extensive experiments on a real-world dataset from Meituan show that our method achieves over 50% higher accuracy in predicting Route RSRP compared to state-of-the-art baselines.

Piezoelectric nanopositioning systems are often limited by lightly damped structural resonances and the gain–phase constraints of linear feedback, which restrict achievable bandwidth and tracking performance. This paper presents a dual-loop architecture that combines an inner-loop non-minimum-phase resonant controller (NRC) for active damping with an outer-loop tracking controller augmented by a constant-gain, lead-in-phase (CgLp) reset element to provide phase lead at the targeted crossover without increasing loop gain. We show that aggressively tuned CgLp designs with larger phase lead can introduce pronounced higher-order harmonics, degrading error sensitivity in specific frequency bands and causing multiple-reset behavior. To address this, a shaping filter is introduced in the reset-trigger path to regulate the reset action and suppress harmonic-induced effects while preserving the desired crossover-phase recovery. The proposed controllers are implemented in real time on an industrial piezo nanopositioner, demonstrating an experimental open-loop crossover increase of approximately 55~Hz and a closed-loop bandwidth improvement of about 34~Hz relative to a well-tuned linear baseline.

With generative AI advancing, empathy in human-AI interaction is essential. While prior work focuses on emotional reflection, emotional exploration, key to deeper engagement, remains overlooked. Existing LLMs rely on text which captures limited emotion nuances. To address this, we propose RE-LLM, a speech-LLM integrating dimensional emotion embeddings and auxiliary learning. Experiments show statistically significant gains in empathy metrics across three datasets. RE-LLM relatively improves the Emotional Reaction score by 14.79% and 6.76% compared to text-only and speech-LLM baselines on ESD. Notably, it raises the Exploration score by 35.42% and 3.91% on IEMOCAP, 139.28% and 9.83% on ESD, and 60.95% and 22.64% on MSP-PODCAST. It also boosts unweighted accuracy by 5.4% on IEMOCAP, 2.3% on ESD, and 6.9% on MSP-PODCAST in speech emotion recognition. These results highlight the enriched emotional understanding and improved empathetic response generation of RE-LLM.

Speech Enhancement (SE) in audio devices is often supported by auxiliary modules for Voice Activity Detection (VAD), SNR estimation, or Acoustic Scene Classification to ensure robust context-aware behavior and seamless user experience. Just like SE, these tasks often employ deep learning; however, deploying additional models on-device is computationally impractical, whereas cloud-based inference would introduce additional latency and compromise privacy. Prior work on SE employed Dynamic Channel Pruning (DynCP) to reduce computation by adaptively disabling specific channels based on the current input. In this work, we investigate whether useful signal properties can be estimated from these internal pruning masks, thus removing the need for separate models. We show that simple, interpretable predictors achieve up to 93% accuracy on VAD, 84% on noise classification, and an R2 of 0.86 on F0 estimation. With binary masks, predictions reduce to weighted sums, inducing negligible overhead. Our contribution is twofold: on one hand, we examine the emergent behavior of DynCP models through the lens of downstream prediction tasks, to reveal what they are learning; on the other, we repurpose and re-propose DynCP as a holistic solution for efficient SE and simultaneous estimation of signal properties.

Due to recent advancements in Large Audio-Language Models (LALMs) that demonstrate remarkable performance across a range of sound-, speech- and music-related tasks, there is a growing interest in proposing benchmarks to assess these models. Existing benchmarks generally focus only on reasoning with internal knowledge, neglecting real-world scenarios that require external information grounding. To bridge this gap, we introduce AudioRAG, a novel benchmark designed to evaluate audio-based reasoning augmented by information retrieval in realistic web environments. This benchmark comprises both LLM-generated and manually curated question-answer pairs. Our evaluations reveal that even the state-of-the-art LALMs struggle to answer these questions. We therefore propose an agentic pipeline that integrates audio reasoning with retrieval-augmented generation, providing a stronger baseline for future research.

Packet-level discrete-event simulation (PLDES) is a prevalent tool for evaluating detailed performance of large model training. Although PLDES offers high fidelity and generality, its slow performance has plagued networking practitioners. Existing optimization techniques either simplify the network model, resulting in large errors; or execute it in parallel using multiple processors, with an upper bound on speedup. This paper explores an alternative optimization direction that reduces the computational loads of PLDES while maintaining high fidelity. Our key insight is that, in distributed LLM training, packet-level traffic behaviors often exhibit repetitive contention patterns and steady-states where flow rates stabilize, ignoring these redundant discrete events speeds up the simulation considerably and the error is negligible. We realize this idea by proposing Wormhole, a user-transparent PLDES kernel capable of automatically memoization for unsteady-states and skipping for steady-states. Wormhole adopts network partitioning, state memoization and reuse, and rate-based steady-state identification to accurately determine the periods of each flow's steady-state, while maintaining simulation consistency after fast-forwarding. Experiments demonstrate that Wormhole can achieve a 744x speedup over the original ns-3 (510x for MoE workload), with a bounded error of <1%. Applying current multithreading parallel techniques and Wormhole together allows a 1012x speedup, reducing the simulation time for one GPT-13B training under 128 GPUs from 9 hours to 5 minutes.

Underwater Image Enhancement (UIE) is an ill-posed problem where natural clean references are not available, and the degradation levels vary significantly across semantic regions. Existing UIE methods treat images with a single global model and ignore the inconsistent degradation of different scene components. This oversight leads to significant color distortions and loss of fine details in heterogeneous underwater scenes, especially where degradation varies significantly across different image regions. Therefore, we propose SUCode (Semantic-aware Underwater Codebook Network), which achieves adaptive UIE from semantic-aware discrete codebook representation. Compared with one-shot codebook-based methods, SUCode exploits semantic-aware, pixel-level codebook representation tailored to heterogeneous underwater degradation. A three-stage training paradigm is employed to represent raw underwater image features to avoid pseudo ground-truth contamination. Gated Channel Attention Module (GCAM) and Frequency-Aware Feature Fusion (FAFF) jointly integrate channel and frequency cues for faithful color restoration and texture recovery. Extensive experiments on multiple benchmarks demonstrate that SUCode achieves state-of-the-art performance, outperforming recent UIE methods on both reference and no-reference metrics. The code will be made public available at https://github.com/oucailab/SUCode.

Motivated by active wing flutter suppression in high-Mach-number flight, this paper presents a rapid boundary stabilization strategy for a two-dimensional PDE-modeled elastic plate with in-domain instabilities, where the exponential stability is achieved with a decay rate that can be arbitrarily assigned by the users. First, the aeroelastic system is modeled as two-dimensional coupled wave PDEs with internal anti-damping terms, derived by Piston theory and Hamilton's principle. Using Fourier series expansion, the 2-D problem is decomposed into a large-scale 1-D system, based on which full-state boundary feedback control is designed via PDE backstepping transformation. To enable output-feedback implementation, a state observer is further designed to estimate the distributed states over the two-dimensional spatial domain using the available boundary measurements. Through Lyapunov analysis, the exponential stability of the 2-D elastic plate PDE under the proposed boundary control is established with a designer-tunable decay rate. Numerical simulations verify the effectiveness of the control strategy in suppressing flow-induced vibrations in a 2-D elastic plate.