Recent neural audio compression models often rely on residual vector quantization for high-fidelity coding, but using a fixed number of per-frame codebooks is suboptimal for the wide variability of audio content-especially for signals that are either very simple or highly complex. To address this limitation, we propose SwitchCodec, a neural audio codec based on Residual Experts Vector Quantization (REVQ). REVQ combines a shared quantizer with dynamically routed expert quantizers that are activated according to the input audio, decoupling bitrate from codebook capacity and improving compression efficiency. This design ensures full training and utilization of each quantizer. In addition, a variable-bitrate mechanism adjusts the number of active expert quantizers at inference, enabling multi-bitrate operation without retraining. Experiments demonstrate that SwitchCodec surpasses existing baselines on both objective metrics and subjective listening tests.

In real-world robotic manipulation, states typically admit a neighborhood of near-equivalent actions. That is for each state, there exist a feasible action neighborhood (FAN) rather than a single correct action, within which motions yield indistinguishable progress. However, prevalent VLA training methodologies are directly inherited from linguistic settings and does not exploit the FAN property, thus lead to poor generalization and low sample efficiency. To address this limitation, we introduce a FAN-guided regularizer that shapes the model's output distribution to align with the geometry of FAN. Concretely, we introduce a Gaussian prior that promotes locally smooth and unimodal predictions around the preferred direction and magnitude. In extensive experiments across both reinforced finetuning (RFT) and supervised finetuning (SFT), our method achieves significant improvement in sample efficiency, and success rate in both in-distribution and out-of-distribution (OOD) scenarios. By aligning with the intrinsic action tolerance of physical manipulation, FAN-guided regularization provides a principled and practical method for sample-efficient, and generalizable VLA adaptation. Code is provided in supplemental material.

Navigating to out-of-sight targets from human instructions in unfamiliar environments is a core capability for service robots. Despite substantial progress, most approaches underutilize reusable, persistent memory, constraining performance in lifelong settings. Many are additionally limited to single-modality inputs and employ myopic greedy policies, which often induce inefficient back-and-forth maneuvers (BFMs). To address such limitations, we introduce SSMG-Nav, a framework for object navigation built on a Semantic Skeleton Memory Graph (SSMG) that consolidates past observations into a spatially aligned, persistent memory anchored by topological keypoints (e.g., junctions, room centers). SSMG clusters nearby entities into subgraphs, unifying entity- and space-level semantics to yield a compact set of candidate destinations. To support multimodal targets (images, objects, and text), we integrate a vision-language model (VLM). For each subgraph, a multimodal prompt synthesized from memory guides the VLM to infer a target belief over destinations. A long-horizon planner then trades off this belief against traversability costs to produce a visit sequence that minimizes expected path length, thereby reducing backtracking. Extensive experiments on challenging lifelong benchmarks and standard ObjectNav benchmarks demonstrate that, compared to strong baselines, our method achieves higher success rates and greater path efficiency, validating the effectiveness of SSMG-Nav.

Fully Test Time Adaptation (FTTA) is usually improved through test time objectives and update rules, whereas the suitability of the source-trained representation hierarchy for downstream adaptation remains underexplored. We observe that under corruption, useful predictive cues are not uniformly lost across depth; instead, they tend to weaken progressively along the feature hierarchy. This suggests that standard discriminative source training may produce representations that are effective for clean recognition but not necessarily well suited for later adaptation under distribution shift. Motivated by this observation, we propose Reconstruction-Regularized Source Training for downstream FTTA. During source training, lightweight reconstruction heads are attached to selected intermediate features to encourage the network to retain more input-related structure. These auxiliary heads are discarded after training, so the deployed model keeps the same inference path and can be combined with existing FTTA methods without modifying their test time objectives, update rules, or inference procedure. Experiments on CIFAR10-C, CIFAR100-C, and ImageNet-C across multiple architectures and downstream FTTA methods show a consistent high-level pattern: improving source side training improves the average final performance of the overall FTTA pipeline. Additional analyses further show that the gain consistently includes a stronger corrupted-source starting point and, in some settings, more favorable adaptation dynamics.

Optimal transport (OT) has recently been shown as a promising criterion for unsupervised restoration when no explicit prior model is available. Despite its theoretical appeal, OT still significantly falls short of supervised methods on challenging tasks such as super-resolution, deraining, and dehazing. In this paper, we propose a sparsity-aware optimal transport (SOT) framework to bridge this gap by leveraging a key observation: the degradations in these tasks exhibit distinct sparsity in the frequency domain. Incorporating this sparsity prior into OT can significantly reduce the ambiguity of the inverse mapping for restoration and substantially boost performance. We provide analysis to show exploiting degradation sparsity benefits unsupervised restoration learning. Extensive experiments on real-world super-resolution, deraining, and dehazing demonstrate that SOT offers notable performance gains over standard OT, while achieving superior perceptual quality compared to existing supervised and unsupervised methods. In particular, SOT consistently outperforms existing unsupervised methods across all three tasks and narrows the performance gap to supervised counterparts.

Test-time adaptation (TTA) aims to enhance the cross-domain performance of pre-trained models by adapting to unlabeled test data.While most existing TTA methods rely on backpropagation (BP) for finetuning, BP-free methods such as zeroth-order (ZO) methods are more desired in practical on-device scenarios. ZO methods rely only on forward computation, which can largely reduce the complexity and memory overhead of on-device deployment.However, ZO methods suffer from much higher variance compared with first-order methods in estimating the gradient.To address this, we propose an improved ZO method to substantially boost the performance of ZO optimization based TTA.First, we provide an observation to reveal the persistent low-rank Hessian structure of the loss during the adaptation process. Based on this insight, we then propose a loss-landscape curvature-aware zeroth-order (CAZO) method, which leverages a sliding-average estimation of the diagonal Hessian to construct a covariance matrix for anisotropic‌ perturbation sampling. CAZO operates by freezing pretrained weights and optimizing minimal adapter parameters via forward-only passes based gradient estimation, which can substantially reduce the memory overhead compared to BP-based methods. Extensive experiments demonstrate that CAZO significantly outperforms existing TTA methods, achieving state-of-the-art performance while maintaining an excellent balance between accuracy and memory efficiency. Code is provided in supplemental material.

The human brain is a biologically instantiated on-device neural system that integrates both learning and inference in a unified architecture, which enables rapid and flexible learning on-the-fly. This extraordinary capability is achieved through non-BP learning mechanisms, whereas BP is computationally and memory intensive that unsuitable for on-chip edge learning. Zeroth-order (ZO) optimization methods, which resemble biologically plausible perturbation-based learning, offer a promising alternative that enables learning with only forward passes and hence can significantly reduce the complexity of on-chip hardware implementation. However, in this work we show that applying ZO methods to spiking neural networks (SNNs) is non-trivial due to the step-function nature of spiking activation. We analyze the challenges posed by the spiking activation, and reveal a variance amplification effect of it. Based on this insight, we propose a subspace-based ZO (SZO) method that leverages the intrinsic low-dimensional structure of the SNN optimization trajectory. By learning in a low-dimensional subspace, SZO substantially enhances ZO learning efficacy, achieving accuracy comparable to first-order (FO) methods with faster learning speed than full-space BP. We evaluate SZO on model training from scratch, continual training, and unsupervised adaptation. Experimental results demonstrate that SZO closely approaches FO training performance for the first time while offering fast learning speed.

The hardware convergence of Artificial Neural Networks (ANNs) and Spiking Neural Networks (SNNs) is hindered by conflicting computational paradigms: dense tensor parallelism versus asynchronous sparse dynamics. Existing unifications typically rely on inefficient spatial partitioning or mode-reconfigurable datapaths, limiting the flexibility needed by heterogeneous ANN-SNN hybrid models requiring frequent cross-domain interaction. To resolve this, we present NeuroUNI, a unified event-driven multi-core architecture. Unlike partitioned designs, NeuroUNI unifies computation at the primitive level using a novel Five-Tuple Event Model, abstracting both continuous activations and discrete spikes to naturally leverage dynamic sparsity. The architecture features a co-optimized hierarchical Macro-Micro-mu OP ISA, a superscalar SIMD-based microarchitecture, and a decentralized multi-core synchronization protocol. Validated in TSMC 28nm technology via post-synthesis simulation and on a Xilinx VCU129 FPGA prototype, NeuroUNI demonstrates competitive cross-paradigm efficiency. It achieves 35.0x and 1.21x the ANN energy efficiency of the NVIDIA V100 and EyerissV2, respectively, while delivering 5.7x the SNN throughput of TrueNorth. In a unified mapless navigation workload, NeuroUNI attains 422.6 GOPS/W (ANN) and 190.8GSOPS/W (SNN), outperforming Loihi 1 with 56.7x the throughput and 3.85x the energy efficiency, proving the viability of a primitive/ISA-level unified silicon substrate.

Robust 3D registration is a fundamental problem in computer vision and robotics, where the goal is to estimate the geometric transformation between two sets of measurements in the presence of noise and outlier contamination. Existing robust registration methods are mainly built on either maximum consensus (MC) estimators, which first identify inliers and then estimate the transformation, or M-estimators, which directly optimize a robust objective. However, MC-based methods typically ignore residual magnitudes during inlier selection, while many M-estimators do not explicitly couple inlier/outlier identification with model estimation. Thus, a unified and efficient framework that jointly performs inlier identification and accurate transformation estimation remains desirable for challenging 3D registration. In this work, we introduce a unified truncated-loss based formulation for simultaneous inlier identification and model estimation (SIME) and study it in the context of 3D registration. We show that, compared with MC-based robust fitting, SIME can achieve a lower fitting residual because it incorporates residual magnitudes into the inlier selection process. To solve the resulting nonconvex problem, we develop an alternating minimization (AM) algorithm, and further propose an AM method embedded with semidefinite relaxation (AM-R) to alleviate the difficulty caused by the binary inlier variables. We instantiate the proposed framework for 3D rotation search and rigid point-set registration using quaternion-based formulations. Experimental results on both simulated and real-world registration tasks demonstrate that the proposed methods compare favorably with strong baseline solvers, especially in high noise and extreme outliers. In the synthetic experiments, the proposed methods are evaluated under outlier ratios up to 95% and consistently achieve competitive or better accuracy, with clear advantages in high-noise cases. On 3DMatch, SIME (AM) achieves a mean registration success rate of 91.0%. These results show the potential of SIME for reliable 3D registration in practical robotics, computer vision, and geometric perception applications.

Generative speech enhancement (SE) methods, typically implemented with diffusion or flow matching, exhibit strong generalization to unseen acoustic conditions. However, they often underperform discriminative approaches in matched, in-domain settings. Recent attempts to close this gap have relied on pretrained discriminative models or cascaded multi-flow architectures, which add inference steps and increase computational cost. To address these limitations, we propose HyFlowSE, a flow matching SE framework with hybrid generative–discriminative learning. HyFlowSE leverages neural ordinary differential equations (ODEs) for end-to-end training and jointly optimizes generative and discriminative objectives within a single model. Experiments on popular benchmark datasets show that HyFlowSE, with only 5.2 M parameters, outperforms other generative SE methods across nearly all evaluation metrics, with especially pronounced gains at low signal-to-noise ratios.

Plane segmentation algorithms are widely used in robotics, serving key roles in scenarios such as indoor localization, scene understanding, and robotic manipulation. These applications typically require real-time, precise, and robust plane segmentation processing, which presents a significant challenge. Existing methods based on pixel-wise or fix-sized patch-wise operation is redundant, as planar regions in real-world scenes are of diverse sizes. In this paper, we introduce a highly efficient method for plane segmentation, namely Adaptive Patch-wise Region Growing (APRG). APRG begins with data sampling to construct a data pyramid. To avoid redundant planer fitting in large planar regions, we introduce an adaptive patch-wise plane fitting algorithm with the pyramid accessed in a top-down manner. The largest possible planar patches are obtained in this process. Subsequently we introduce a region growing algorithm specially designed for our patch representation. Overall, APRG achieves more than 600 FPS at a 640x480 resolution on a mid-range CPU without using parallel acceleration techniques, which outperforms the state-of-the-art method by a factor of 1.46. Besides, in addition to its speedup in run-time, APRG significantly improves the segmentation quality, especially on real-world data.

The majority of deep learning-based speech enhancement techniques rely on supervised training, which requires extensive pairs of noisy and clean speech. However, due to the complexity and variability of real-world environments, obtaining ground-truth clean speech can be challenging or even impractical in certain situations. To tackle this issue, we introduce a novel self-supervised speech enhancement method that eliminates the need for clean reference speech. Our method involves two training stages to develop a speech enhancement model through iterative refinement. In the first stage, we use unprocessed noisy speech and noise to create noisier-to-noisy data pairs, which are used to train the initial model. In the second stage, we iteratively generate noisier-to-noisy data pairs using speech sampled from the estimated speech and unprocessed noisy speech, along with noise sampled from the noise corpus, to train a more effective model. Further, we provide analysis to compare and deepen the understanding of various self-supervised learning methods, including NyTT, IDR-SE, and the proposed SelfSE method. Particularly, we show that noisy-target based self-supervised learning inherently introduces a bias, and this bias is SNR-dependent that it increases as the SNR decreases. We conduct extensive experiments on three popular benchmark datasets, and the results demonstrate that our approach achieves performance comparable to supervised learning methods on simulated data and surpasses them on real-world data in both speech enhancement and recognition tasks.

Recently, learning-based robotic navigation systems have gained extensive research attention and made significant progress. However, the diversity of open-world scenarios poses a major challenge for the generalization of such systems to practical scenarios. Specifically, learned systems for scene measurement and state estimation tend to degrade when the application scenarios deviate from the training data, resulting in unreliable depth and pose estimation. Toward addressing this problem, this work aims to develop a visual odometry (VO) system that can rapidly adapt to diverse novel environments in an online manner. To this end, we construct a self-supervised online adaptation framework where the VO system and the depth estimation module reinforce each other through a continuous measure-calibrate-remeasure cycle. First, we design a monocular depth estimation network with lightweight refiner modules, which enables efficient online adaptation. Then, we construct an objective for self-supervised learning of the depth estimation module based on the output of the VO system and the contextual semantic information of the scene. Specifically, a sparse depth densification (SDD) module and a dynamic consistency enhancement (DCE) module are proposed to leverage camera poses and contextual semantics to generate pseudo-depths and valid masks for the online adaptation. Finally, we demonstrate the robustness and generalization capability of the proposed method in comparison with state-of-the-art learning-based approaches on urban, in-house datasets and a robot platform. Code is publicly available at: https://github.com/jixingwu/SOL-SLAM

Test-time adaptation (TTA) aims to adapt a source model to a target domain using only test data. Existing methods predominantly rely on unsupervised entropy minimization or its variants, which suffer from degeneration, leading to trivial solutions with low-entropy but inaccurate predictions. In this work, we identify *entropy-deceptive* (ED) samples, instances where the model makes highly confident yet incorrect predictions, as the underlying cause of degeneration. Further, we reveal that the gradients of entropy minimization in TTA have an intrinsic low-dimensional structure, driven primarily by *entropy-truthful* (ET) samples whose gradients are highly correlated. In contrast, ED samples have scattered, less correlated gradients. Leveraging this observation, we show that the detrimental impact of ED samples can be suppressed by constraining model updates within the principal subspace of backward gradients. Building on this insight, we propose LCoTTA, a lifelong continual TTA method that tracks the principal subspace of gradients online and utilizes their projections onto this subspace for adaptation. Further, we provide theoretical analysis to show that the proposed subspace-based method can enhance the robustness against detrimental ED samples. Extensive experiments demonstrate that LCoTTA effectively overcomes degeneration and significantly outperforms existing methods in long-term continual adaptation scenarios. Code is available online.

Semi-gradient Q-learning is widely applied across various fields; however, due to the absence of an explicit loss function, understanding its preference for critical points in the parameter space-namely, the implicit bias-remains challenging. This paper leverages the Fokker-Planck equation to construct and visualize the non-equilibrium loss landscape in a two-dimensional parameter space. Our visualization reveals that the semi-gradient method transforms stable critical points into unstable ones, resulting in training dynamics biased against unstable critical points. Furthermore, we show that this phenomenon extends to high-dimensional parameter spaces and neural network settings, such as DQN. This paper provides valuable insights into the implicit bias of semi-gradient Q-learning.

Fine-tuning pre-trained large language models (LLMs) on downstream tasks has achieved significant success across various domains. However, as model sizes grow, traditional first-order fine-tuning algorithms incur substantial memory overhead due to the need for activation storage for back-propagation (BP). The BP-free Memory-Efficient Zeroth-Order Optimization (MeZO) method estimates gradients through finite differences, avoiding the storage of activation values, and has been demonstrated as a viable approach for fine-tuning large language models. This work proposes the Multiple-query Memory Efficient Zeroth-Order (MUZO) method, which is based on variance-reduced multiple queries to obtain the average of gradient estimates. When combined with Adam optimizer, MUZO-Adam demonstrates superior performance in fine-tuning various LLMs. Furthermore, we provide theoretical guarantees for the convergence of the MUZO-Adam optimizer. Extensive experiments empirically demonstrate that MUZO-Adam converges better than MeZO-SGD and achieves near first-order optimizer performance on downstream classification, multiple-choice, and generation tasks.

Synaptic delay has attracted significant attention in neural network dynamics for integrating and processing complex spatiotemporal information. This paper introduces a high-throughput Spiking Neural Network (SNN) processor that supports synaptic delay-based emulation for edge applications. The processor leverages a multicore pipelined architecture with parallel compute engines, capable of real-time processing of the computational load associated with synaptic delays. We develop a SoC prototype of the proposed processor on PYNQ Z2 FPGA platform and evaluate its performance using the Spiking Heidelberg Digits (SHD) benchmark for low-power keyword spotting tasks. The processor achieves 93.4% accuracy in deployment and an average throughput of 104 samples/sec at a typical operating frequency of 125 MHz and 282 mW power consumption.

Fully test-time adaptation (TTA) aims to address the cross-domain performance degradation problem by adapting a source pre-trained model on test data. Ensemble methods have been shown to be very effective for out-of-distribution (OOD) uncertainty estimation, and are promising for TTA. However, naive ensembling of multiple models would significantly increase the overall model size and computational load. This work explores more efficient ensemble approaches for TTA that can improve the cross-domain performance but without significant increase of the overall model size and computational load. To this end, our contribution is twofold. First, we show that, compared with using a single model, an ensemble of multiple smaller models with fewer or comparable overall parameters can also markedly improve TTA performance. Second, we propose a novel self-ensemble method that ensembling is performed on multiple intermediate outputs of a single model rather than on multiple independent models. The proposed method only incurs slight increase in parameters, and can be readily incorporated into existing TTA methods in a plug-in manner. Experimental results on benchmark datasets demonstrate that the proposed self-ensemble method can significantly improve the performance of existing state-of-the-art TTA methods. Furthermore, when combining self-ensemble with the ensembling of smaller models, even greater improvement can be achieved. Code is available at: https://github.com/BryantLeedy/SelfEnsembleTTA.

Model reparameterization is a widely accepted technique for improving inference speed without compromising performance. However, current Post-training Quantization (PTQ) methods often lead to significant accuracy degradation when applied to reparameterized models. This is primarily caused by channel-specific and sample-specific outliers, which appear only at specific samples and channels and impact on the selection of quantization parameters. To address this issue, we propose RepAPQ, a novel framework that preserves the accuracy of quantized reparameterization models. Different from previous frameworks using Mean Squared Error (MSE) as a measurement, we utilize Mean Absolute Error (MAE) to mitigate the influence of outliers on quantization parameters. Our framework comprises two main components: Quantization Protecting Reparameterization and Across-block Calibration. For effective calibration, Quantization Protecting Reparameterization combines multiple branches into a single convolution with an affine layer. During training, the affine layer accelerates convergence and amplifies the output of the convolution to better accommodate samples with outliers. Additionally, Across-block Calibration leverages the measurement of stage output as supervision to address the gradient problem introduced by MAE and enhance the interlayer correlation with quantization parameters. Comprehensive experiments demonstrate the effectiveness of RepAPQ across various models and tasks. Our framework outperforms previous methods by approximately 1\% for 8-bit PTQ and 2\% for 6-bit PTQ, showcasing its superior performance. The code is available at \url{https://github.com/ilur98/DLMC-QUANT}.

Deep learning-based speech enhancement methods are predominantly trained in a supervised manner, relying on synthesized paired noisy-to-clean data. However, acquiring clean speech in real-world scenarios is often difficult or even impractical. To overcome this limitation, we propose a novel unsupervised learning framework for speech enhancement that relies solely on observed noisy speech, called MOS-GAN. Specifically, we leverage generative adversarial networks (GANs), where the generator (the enhancement model) is optimized to maximize the mean opinion score (MOS) guided by a discriminator, while the discriminator (a non-intrusive speech quality metric model) is optimized to predict MOS. However, without using reference clean speech, directly training of MOS-GAN is unstable and cannot achieve satisfactory performance. To address this issue, we further incorporate an unsupervised prior loss to substantially enhance training performance. Experimental results on benchmarks demonstrate that the proposed method, which requires neither clean data nor teacher models, performs on par with leading self-supervised and unsupervised approaches.