Searcharxiv⌕ Search

arXiv subjects

Xinjue Wang

Publications and source records attributed to Xinjue Wang.

12 recordsLinked to original sources

Zeroth-Order Riemannian Optimization on Fixed-Rank Manifolds for LLM Fine-Tuning

Zeroth-order (ZO) optimization estimates descent directions from function values alone, which removes the activation storage required by backpropagation and makes it attractive for memory-efficient fine-tuning of large language models (LLMs). Existing low-rank ZO methods either parameterize the additive update through Euclidean factors with a non-unique representation, or restrict each finite-difference direction to be low-rank. In this work, we propose to constrain each additive update block to a fixed-rank matrix manifold while keeping the pretrained weights frozen, so the learned update remains rank-constrained throughout fine-tuning. On this geometry, we build a Riemannian ZO gradient estimator that projects a random matrix onto the tangent space, normalizes it to correct the scale mismatch induced by projection, and further propose Zeroth-order Riemannian Gradient Descent (ZO-RGD) and its smoothed variant, ZO-RGD-S. We show that the ambient finite difference matches its retracted counterpart, the normalized estimator recovers the Riemannian gradient in expectation, and ZO-RGD converges under trajectory-regularity conditions. Experiments on LLM fine-tuning tasks show that the proposed algorithms are competitive with strong forward-only baselines across model families and tasks, with gains in several settings.

cs.LG↗

HC2Robot: Hyper-Connection Transformers with Predicted Foresight for Multi-Terrain Humanoid Parkour

Humanoid parkour requires robots to reason about future footholds and whole-body motions over highly discontinuous terrain. We revisit the information asymmetry in asymmetric actor-critic learning for perceptive humanoid locomotion: the critic exploits privileged simulator signals during training, whereas the deployable actor operates solely from onboard perception. We argue that the key is not exposing privileged information to the actor, but learning a deployable approximation of it to narrow the information gap between training and deployment. We present HC2Robot, a deployable perceptive humanoid parkour framework consisting of a Predicted Foresight Module (PFM) and a Hyper-Connection (HC) Transformer. PFM predicts from deployable observations the policy-conditioned timing, location, and support quality of each foot's next touchdown event, converting privileged supervision into a deployable latent representation that provides the actor with predicted touchdown information without privileged runtime sensing. The HC Transformer further enhances temporal reasoning through learnable multi-stream routing, while a future motion prediction head provides short-horizon motion priors for regularization. Across stairs, gaps, high platforms, and slopes, HC2Robot achieves a mean simulated traversal success rate of 97.60% and consistently higher safe foothold rates than strong perceptive locomotion baselines, supporting improved traversal performance and geometric foothold support under deployable sensing.

cs.RO↗

Efficient SCA Algorithms for Sparse Bayesian Learning

Type-II sparse Bayesian learning (SBL) for the multiple measurement vector model recovers a sparse signal power vector by minimizing the marginal negative log-likelihood via the expectation-maximization (EM) algorithm, referred to as M-SBL. We revisit the covariance learning (CL) problem and develop three parallel first-order algorithms within the successive convex approximation (SCA) framework. CL-SCA minimizes per-coordinate surrogates formed by partially linearizing the log-determinant term, while CL-FP uses a damped projected fixed-point iteration derived from stationarity conditions. Both methods have closed-form updates. We prove that the fixed points of both update maps are exactly the KKT points of the CL problem, that each update direction satisfies a quantitative descent inequality, that the damped iterates achieve sufficient decrease with an $\mathcal{O}(T^{-1/2})$ best-iterate stationarity rate, and we derive a backtracking rule that retains this rate without step-size tuning. If damping factors are bounded away from zero, the residual vanishes along the whole sequence and every limit point is a KKT point. Local analysis motivates CL-Hyb, which combines CL-SCA and CL-FP. Experiments on device activity detection in massive random access show that CL-Hyb-bt matches the best CL solvers' accuracy at a fraction of the runtime and is over an order of magnitude faster than EM.

eess.SP↗

ParkourFormer: Integrating Predictive Supervision and Sequence Modeling into Parkour Locomotion

Humanoid parkour requires locomotion policies to coordinate whole-body dynamics across rapidly changing terrains such as stairs, gaps, slopes, and obstacles. Existing learning-based humanoid locomotion policies typically infer actions from current and historical observations, while the future evolution of the robot state is represented only implicitly. Explicitly anticipating short-horizon state evolution is particularly beneficial for multiple terrains locomotion, where current actions must account for upcoming contact transitions and whole-body dynamics. We present ParkourForme, a Transformer-based sequence modeling framework that reformulates humanoid locomotion as a future-conditioned decision-making problem. The current robot state queries historical sensorimotor trajectories through cross-attention, while a lightweight prediction head forecasts short-horizon future proprioceptive states. The future states are predicted solely from historical observations and actions, with ground-truth future states used only as supervision targets, and are fused with temporal features to generate the current action without future-information leakage. We evaluate ParkourFormer on a diverse multi-terrain humanoid parkour benchmark including stairs, gaps, slopes, rough terrain, and obstacle traversal. ParkourFormer achieves a 93.96% mean traversal success rate, compared with 90.49% for a vanilla Transformer and 87.16% for a MoE-MLP policy, while maintaining a single policy across all terrain types. These results demonstrate that explicit future-state modeling significantly improves robustness and generalization for agile whole-body locomotion.

cs.RO↗

A Majorization-Minimization Framework for Activity Detection in Mixed Near-and-Far-Field Random Access

Grant-free random access in massive machine-type communications requires detecting a small active set from length-L uplink pilots received at an M-antenna base station. Classical covariance-based detectors are largely built on the far-field (FF) model, where the M antenna-domain observations are treated as independent snapshots, and inference reduces to an L $\times$ L covariance description. This model becomes inadequate in mixed near-field (NF) and FF access, where NF devices induce device-specific structured spatial covariances and the aggregate observation no longer fits the FF snapshot structure. In this paper, we develop a unified covariance-aware Rician framework for activity detection that treats FF and NF devices by a single likelihood model. Within this framework, we propose a majorization-minimization projected gradient descent (MM-PGD) detector for the resulting relaxed likelihood. For scalable exact likelihood evaluation, we further derive a Kronecker-Woodbury implementation that exploits the mixed NF/FF covariance structure, and avoids factorizing the full LM $\times$ LM covariance matrix. Numerical results show that MM-PGD is on par with the strongest coordinate-wise baseline in the all-FF limit, while its advantage becomes more pronounced as the fraction of NF users increases under the tested regimes.

eess.SP↗

Covariance-Aware MM-PGD for Mixed Near-/Far-Field Activity Detection

Grant-free activity detection with mixed near-field (NF) and far-field (FF) devices is an important problem that can be addressed via covariance-based detectors. The difficulty is that NF users induce device-specific structured spatial covariances, whereas FF users are well approximated by isotropic covariances. Under a unified Rician model, we first formulate activity detection as a relaxed maximum-likelihood problem in the full LM-dimensional vectorized observation space. We then develop a covariance-aware majorization-minimization projected gradient descent (MM-PGD) detector. It updates the full activity vector jointly and avoids the per-coordinate high-rank subproblems that arise in the NF regime. Numerical results over SNR, antenna-count, and NF-ratio sweeps show that MM-PGD achieves up to 20x lower miss-detection probability than the strongest coordinate-wise baseline. The advantage is most pronounced at high NF ratios, while in the all-FF case MM-PGD performs on par with the strongest coordinate-wise baseline.

eess.SP↗

GeoMix: Descriptor-Free Visual Localization via Global Context and Multi-Detector Training

Descriptor-free visual localization eliminates high-dimensional descriptor storage, preserves scene privacy, and simplifies map maintenance, yet its accuracy still lags far behind descriptor-based pipelines. We identify this gap to insufficient geometric discriminability in geometry-only matching. Without visual appearance, current methods underutilize local geometry cues, lack the global context among keypoints, and overfit to a single keypoint detector. We further observe that descriptor-free matching naturally enables multi-detector training, as heterogeneous keypoints can be optimized in a shared geometry-only space without aligning descriptor spaces. Building on these insights, we propose GeoMix, a descriptor-free 2D-3D matching framework that strengthens geometric discriminability at three levels. Locally, directional and distance-aware embeddings enrich neighborhood aggregation with fine-grained spatial structure. Globally, learnable context nodes aggregate and redistribute scene-wide information via cross-attention to resolve ambiguities beyond local receptive fields. At the training level, Mix-Training exploits this detector-agnostic geometry space to learn representations across multiple keypoint detectors. Extensive experiments on MegaDepth, Cambridge Landmarks, 7Scenes, and Aachen Day-Night show that GeoMix sets a new state of the art among descriptor-free methods, reducing 75th-percentile rotation error by 89\% and translation error by up to 90\% over the previous best, while generalizing zero-shot to unseen detectors and narrowing the gap to descriptor-based pipelines. Code is available at $\href{https://github.com/YejunZhang/Geomix}{\text{this links}}$.

cs.CV↗

Finer Parameter Steps for Low-Rank PEFT: A Controlled Study with CP Tensor Adapters

Low-rank adapters are usually compared by sweeping a small set of ranks, but the rank also fixes the resolution of the parameter budget. For a $2048{\times}2048$ OPT attention projection, increasing LoRA by one rank stores $4096$ trainable scalars, leaving large gaps between feasible low-budget adapter sizes. This paper asks whether a tensorized adapter with finer capacity increments changes the observed accuracy--budget trade-off. We instantiate this question with fixed-component canonical polyadic (CP) tensor adapters. Under a $32{\times}64{\times}32{\times}64$ tensorization, one normalized CP component stores $193$ trainable scalars per projection, about $21$ times smaller than one LoRA rank step. We compare CP adapters and LoRA on OPT-1.3B across SST-2, RTE, and BoolQ under matched target modules, training protocol, data caps, and seed schedules. CP trains stably and fills the gaps between LoRA ranks, but the effect is task-dependent: SST-2 reaches an early low-budget plateau, BoolQ benefits from additional CP components before saturating slightly below LoRA, and RTE remains LoRA-favored. Finer parameter steps are therefore useful for diagnosing PEFT budget sensitivity, but they do not by themselves guarantee a better accuracy--budget curve.

cs.LG↗

Anisotropic Tensor Deconvolution of Hyperspectral Images

Hyperspectral image (HSI) deconvolution is a challenging ill-posed inverse problem, made difficult by the data's high dimensionality.We propose a parameter-parsimonious framework based on a low-rank Canonical Polyadic Decomposition (CPD) of the entire latent HSI $\mathbf{\mathcal{X}} \in \mathbb{R}^{P\times Q \times N}$.This approach recasts the problem from recovering a large-scale image with $PQN$ variables to estimating the CPD factors with $(P+Q+N)R$ variables.This model also enables a structure-aware, anisotropic Total Variation (TV) regularization applied only to the spatial factors, preserving the smooth spectral signatures.An efficient algorithm based on the Proximal Alternating Linearized Minimization (PALM) framework is developed to solve the resulting non-convex optimization problem.Experiments confirm the model's efficiency, showing a numerous parameter reduction of over two orders of magnitude and a compelling trade-off between model compactness and reconstruction accuracy.

eess.IV↗

Robust Activity Detection for Massive Random Access

Massive machine-type communications (mMTC) are fundamental to the Internet of Things (IoT) framework in future wireless networks, involving the connection of a vast number of devices with sporadic transmission patterns. Traditional device activity detection (AD) methods are typically developed for Gaussian noise, but their performance may deteriorate when these conditions are not met, particularly in the presence of heavy-tailed impulsive noise. In this paper, we propose robust statistical techniques for AD that do not rely on the Gaussian assumption and replace the Gaussian loss function with robust loss functions that can effectively mitigate the impact of heavy-tailed noise and outliers. First, we prove that the coordinate-wise (conditional) objective function is geodesically convex and derive a fixed-point (FP) algorithm for minimizing it, along with convergence guarantees. Building on the FP algorithm, we propose two robust algorithms for solving the full (unconditional) objective function: a coordinate-wise optimization algorithm (RCWO) and a greedy covariance learning-based matching pursuit algorithm (RCL-MP). Numerical experiments demonstrate that the proposed methods significantly outperform existing algorithms in scenarios with non-Gaussian noise, achieving higher detection accuracy and robustness.

eess.SP↗

Generalized Nonnegative Structured Kruskal Tensor Regression

This paper introduces Generalized Nonnegative Structured Kruskal Tensor Regression (NS-KTR), a novel tensor regression framework that enhances interpretability and performance through mode-specific hybrid regularization and nonnegativity constraints. Our approach accommodates both linear and logistic regression formulations for diverse response variables while addressing the structural heterogeneity inherent in multidimensional tensor data. We integrate fused LASSO, total variation, and ridge regularizers, each tailored to specific tensor modes, and develop an efficient alternating direction method of multipliers (ADMM) based algorithm for parameter estimation. Comprehensive experiments on synthetic signals and real hyperspectral datasets demonstrate that NS-KTR consistently outperforms conventional tensor regression methods. The framework's ability to preserve distinct structural characteristics across tensor dimensions while ensuring physical interpretability makes it especially suitable for applications in signal processing and hyperspectral image analysis.

eess.SP↗

Graph Convolutional Neural Networks Sensitivity under Probabilistic Error Model

Graph Neural Networks (GNNs), particularly Graph Convolutional Neural Networks (GCNNs), have emerged as pivotal instruments in machine learning and signal processing for processing graph-structured data. This paper proposes an analysis framework to investigate the sensitivity of GCNNs to probabilistic graph perturbations, directly impacting the graph shift operator (GSO). Our study establishes tight expected GSO error bounds, which are explicitly linked to the error model parameters, and reveals a linear relationship between GSO perturbations and the resulting output differences at each layer of GCNNs. This linearity demonstrates that a single-layer GCNN maintains stability under graph edge perturbations, provided that the GSO errors remain bounded, regardless of the perturbation scale. For multilayer GCNNs, the dependency of system's output difference on GSO perturbations is shown to be a recursion of linearity. Finally, we exemplify the framework with the Graph Isomorphism Network (GIN) and Simple Graph Convolution Network (SGCN). Experiments validate our theoretical derivations and the effectiveness of our approach.

stat.ML↗