arXiv subjects

Wei Cao

Publications and source records attributed to Wei Cao.

At least 19 recordsLinked to original sources

arXiv2026

SCOPE: Selective Critical-Step On-Policy Self-Distillation for Agentic Reinforcement Learning

Outcome-based reinforcement learning provides coarse supervision for long-horizon LLM agents: failed trajectories may contain correct steps, while successful trajectories can contain errors. Trajectory-level skills can also shift teacher preferences at unrelated steps, introducing off-target distillation signals. We introduce SCOPE, a framework for Selective Critical-Step On-Policy Self-Distillation. After a supervised warm-up, the policy analyzes its own trajectories, generates step-specific hindsight skills, and distills their guidance only at selected error steps. A bounded signed teacher-student coefficient stabilizes auxiliary updates while preserving positive reinforcement and negative correction. A success-relative pairwise length reward further encourages efficient task completion. Experiments on ALFWorld and WebShop demonstrate consistent improvements across multiple model backbones. With Qwen2.5-3B, SCOPE achieves 97.7% success on ALFWorld and 85.2% on WebShop, outperforming SEED by 5.9 and 6.3 percentage points, respectively. On unseen ALFWorld tasks, SCOPE reaches 89.6% success, improving over GRPO by 23.2 percentage points with 46.7% fewer interactions. Scaling to Qwen3-32B, SCOPE achieves 99.2% success on seen ALFWorld tasks and 94.0% on unseen tasks.

cs.AI↗

arXiv2026

EOPSA: Efficient On-Policy Self-Distilled Safety Alignment

On-Policy Self-Distillation (OPSD) has emerged as a promising paradigm for safety alignment, delivering dense, token-level supervision by distilling from a teacher conditioned on refusal-oriented privileged prompts. However, we reveal that this paradigm suffers from critical inefficiencies that degrade both training efficiency and general reasoning capabilities. Specifically, we diagnose two fundamental bottlenecks: (1) supervisory collapse over extended rollouts, where the teacher's corrective efficacy degrades precipitously as the student's generation prefix lengthens, injecting noisy gradients into late-stage tokens; and (2) gradient dilution from stylistic shifts, where the distillation objective is dominated by safety-irrelevant stylistic discrepancies induced by privileged prompting, washing out genuine safety signals and impairing base reasoning. To resolve these issues, we propose Efficient On-Policy Self-Distilled Safety Alignment (EOPSA), which concentrates computational and gradient budgets exclusively on reliably supervised, safety-critical tokens. EOPSA incorporates two coordinated mechanisms: (i) Adaptive Rollout Scheduling, which dynamically bounds the generation horizon guided by a novel Teacher Rescue Rate (TRR) metric to operate strictly within reliable supervision regimes; and (ii) Selective Distillation, which filters out safety-neutral tokens to restrict gradient updates exclusively to safety-pivotal transitions. Extensive evaluations across reasoning models up to 32B parameters demonstrate that EOPSA slashes rollout computation by $\sim$50% and backpropagates through merely $\sim$2% of tokens, consistently outperforming full-token distillation baselines in both safety compliance and reasoning retention.

cs.AI↗

arXiv2026

Robust Representation Learning for Spiking Neural Networks based on Conditional Information Bottleneck

Spiking Neural Networks (SNNs) have shown great potential for low-power neuromorphic computing due to their event-driven processing and spatiotemporal sparsity, but they remain vulnerable to input perturbations, hardware-intrinsic noise, and adversarial attacks in real-world deployment. We attribute this vulnerability partly to input-specific hidden representations learned by standard SNNs trained with conventional cross-entropy loss, as our information-plane analysis empirically shows that substantial input-related information can still be retained after label-predictive information becomes stable. To address this issue, we propose SCIB, a simple yet effective robust learning objective for SNNs based on the Conditional Information Bottleneck. Different from conventional variational IB methods, SCIB replaces the mutual information compression term $I(X;S)$ with the conditional mutual information $I(X;S|Y)$ and derives a tractable variational upper bound. Moreover, it introduces a learnable class-conditional prior instead of a label-agnostic isotropic Gaussian prior, improving the expressiveness of variational compression and encouraging class-structured representations when combined with the supervised classification objective. Systematic experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 under external environmental noise, neuromorphic hardware noise, and adversarial attacks show that SCIB improves robustness over standard SNNs and IB-based baselines under most evaluated noise conditions, improves adversarial accuracy under the evaluated attacks without using adversarial training, and reduces the SynOps-based single-sample inference energy consumption by 24.9% relative to an unconstrained SNN baseline.

cs.AI↗

arXiv2026

Continual Concept Erasure in Diffusion Models by Suppressing Cross-Edit Interference

Concept erasure removes copyright-protected, privacy-sensitive, or otherwise undesirable concepts from pretrained text-to-image diffusion models to support content governance and compliance. As erasure requests arrive over time, models must remove new targets without undoing prior erasures. Existing methods do not constrain interference across edits: residual perturbations outside the retain set interact and accumulate, degrading unrelated generations and sometimes collapsing previously erased targets into noise. We propose CEASE (Continual Erasure via Adaptive Subspace Editing), a training-free method that imposes two subspace constraints on a closed-form solver. CEASE adds the token representation of the shared replacement to the solver's invariance matrix and, when interference is detected, projects the current update onto the orthogonal complement of dominant output directions extracted from cumulative past updates. A closed-form decomposition attributes the accumulated interference to repeated activation of the shared replacement and overlap between successive update directions, showing that the two constraints suppress these respective sources. Across continual erasure of celebrities, artistic styles, and instances, CEASE achieves the most consistent erase-preserve trade-off, while existing methods either degrade general generation or insufficiently erase targets.

cs.CV↗

arXiv2026

4Director: Controlling Video World Models with Rigid 3D Geometry

Precise control over camera and object motion is essential for professional video production. Existing methods control objects only coarsely, through image-plane cues that are ambiguous in depth and rotation or through 3D tracks and blobs that lack complete geometry and lose consistency across viewpoint changes. We introduce 4Director, a video world model conditioned on an explicit 4D scene representation: each object is reconstructed once from the input image as a canonical mesh and moved by one prescribed rigid transformation per frame. This representation provides an intuitive 3D control interface and prevents unobserved geometry from being regenerated independently in every frame. We render the controlled scene as a depth video and introduce a Motion Adapter that transforms this geometric scaffold into video while synthesizing view-consistent appearance, illumination, and non-rigid dynamics. For training, we construct RealCOD-Rigid, a new dataset of 20,774 clips annotated with rigid 3D scenes by our automatic pipeline. We further introduce Identity-Gated IoU (IG-IoU), which jointly evaluates adherence to prescribed object motion and preservation of object identity. Experiments demonstrate that 4Director consistently outperforms prior methods in visual quality and in camera and object control.

cs.CV↗

arXiv2026

QuantaSpike: Short-Window Spike-Driven Quantization for Large Language Models

Large language models (LLMs) achieve strong performance across many tasks but rely on dense multiply-accumulate (MAC) operations during inference, resulting in high energy cost. Spiking neural networks (SNNs) offer an event-driven alternative in which synaptic integration uses lightweight accumulation. However, spike-driven LLM inference remains difficult because outlier-heavy activations typically require long firing windows or auxiliary non-spiking paths. We propose QuantaSpike, a short-window spike-driven quantization framework for LLMs built around Logarithmic Ternary Integrate-and-Fire (LTIF) neurons. LTIF uses ternary events with power-of-two membrane-response quanta, improving the information represented by each firing step while retaining shift-ACC-compatible computation. QuantaSpike combines this neuron with group-adaptive gain and selective outlier admission: normal values use residual LTIF steps, whereas admitted outliers receive one additional onset spike before entering the same residual dynamics. Across OPT and Llama-2, QuantaSpike achieves state-of-the-art or competitive perplexity and zero-shot accuracy among spike-driven LLM quantization methods. It also transfers to newer dense LLMs, remaining close to the FP16 reference on Llama-3-8B and Qwen3-8B under the same four-step firing window. Analytical linear-energy projections show that QuantaSpike reduces the energy of one linear transformation by about $80.0\%$ on OPT models and $67.1\%$ on Llama-2 models relative to SpikeQuant, providing an accurate and energy-efficient spike-driven path for LLM inference.

cs.AI↗

arXiv2026

Renewable Online Expectile Regression for Heterogeneous Streaming Data with Abnormal Batches

Streaming data, characterized by high volume, rapid arrival rates, and evolving distributions, have become increasingly prevalent in modern applications. Developing efficient and reliable estimation procedures is therefore essential for real-time statistical analysis. However, most existing online estimation methods rely on the assumption of batch homogeneity, which can be violated in practice due to abnormal batches, distributional shifts, or other forms of batch heterogeneity. To address this challenge, we develop renewable online expectile regression procedures for heterogeneous streaming data. Specifically, we propose two complementary strategies for handling abnormal batches: (1) a detection-based approach that employs a sequential monitoring mechanism based on score test statistics to identify and remove potentially abnormal batches; and (2) an adaptive-weighting approach that assigns data-driven weights to incoming batches, reducing the influence of abnormal or drifting batches while retaining information from reliable observations. Both strategies rely solely on score test statistics and can be seamlessly integrated into existing renewable estimation and inference frameworks without requiring additional structural assumptions. Furthermore, to enhance robustness against heavy-tailed errors and outliers, we replace the conventional l2 loss with the Huber loss and develop a robust extension of renew?able online expectile regression. Extensive simulation studies and analyses of clinical datasets demonstrate that the proposed methods achieve improved estimation accuracy and robustness in the presence of batch heterogeneity. Overall, the proposed framework provides a flexible and effective solution for renewable expectile regression in complex streaming data environments.

stat.ME↗

arXiv2026

Programmable and scalable on-chip WDM processing platform enabled by cascaded FSR-free resonators

Wavelength-division multiplexing (WDM) is pivotal for expanding the capacity and functionality of photonic systems, from communications to computing. However, on-chip implementation for broadband operation is fundamentally limited by the narrow free spectral range (FSR) and inefficient drop transmission of conventional optical microcavities. Here, we transcend this limit with a scalable integrated WDM processing platform based on cascaded dual-sided coupled Fabry-Pérot add-drop resonators. This unit achieves a record-large FSR-free bandwidth (> 300 nm) and low insertion loss (< 0.5 dB), while enabling independent manipulation of monochromatic light in both the wavelength and intensity domains. Leveraging these performances, we architect programmable cores with up to 20 channels in a single bus waveguide to demonstrate application versatility by configuring as: a high-capacity on-chip WDM communication fabric (2.4-Tbps single-link WDM transmission and O-to-C-band multi-subband (de)triplexing), a reconfigurable optical processor (broadband and hitless wavelength-selective switching), and a parallel computing accelerator (theoretically capable of 1.28-TOPS convolution operations, 16-channel modulation). This work establishes a unified and versatile WDM processing platform that bridges high-speed optical communication with in-line computing, charting a scalable and backward-compatible path toward petabit-per-second-class links and hundreds of TOPS of on-chip computation, resolving a critical bottleneck for next-generation photonic systems.

physics.optics↗

arXiv2026

MACD: Multi-Agent Clinical Diagnosis with Self-Learned Knowledge for LLM

Large language models (LLMs) have shown promise in supporting medical diagnosis, with prompting-based methods offering a flexible and deployable means of capability enhancement. However, existing prompt engineering and multi-agent approaches often focus on optimizing single inferences, paying less attention to the accumulation of reusable experience from clinical practice, constraining their real-world applicability. To address this, this study proposes a novel Multi-Agent Clinical Diagnosis (MACD) framework, which allows LLMs to self-learn clinical knowledge via a multi-agent pipeline that summarizes, refines, and applies diagnostic insights, mirroring the professional development of human physicians. We further extend it to a MACD-human collaborative workflow, where multiple LLM-based diagnostician agents engage in iterative consultations, supported by a judge agent and human oversight for cases where agreement is not reached. The MIMIC-MACD cohort comprising 4,390 real-world patient cases across seven diseases is constructed, including 1,314 cases for knowledge learning and 3,076 held-out cases for evaluation. Across diverse open-weight LLMs, MACD significantly improves primary diagnostic accuracy, achieving an average improvement of 11.6 percentage points over established authoritative knowledge, while narrowing the performance gap between open-weight models and state-of-the-art LLMs. Furthermore, the MACD-human workflow yields an 18.3-percentage-point improvement over physician-only diagnosis on text-only vignettes, demonstrating the synergistic potential of human-AI collaboration. This work thus presents a scalable self-learning paradigm that bridges the gap between the intrinsic knowledge of LLMs and the demands of real-world clinical practice, advancing towards a reliable, interpretable, and deployable AI-assisted diagnosis.

cs.AI↗

arXiv2026

Polytype-Dependent Upconversion Photoluminescence in 3R-MoS2

Ferroelectric van der Waals materials offer switchable polarization states, yet optical readout of their stacking configurations remains challenging. Here, building on the resonant exciton-exciton annihilation (EEA) mechanism recently identified in 2H-phase TMDs, we report the first observation of upconversion photoluminescence (UPL) in rhombohedral MoS$_2$ and demonstrate that this many-body process is strongly polytype-dependent. Using low-temperature spectroscopy, we observe anti-Stokes emission with superlinear power dependence characteristic of EEA. Beyond serving as an accurate layer-number sensor due to discrete bandgap variations, UPL provides a sensitive probe of stacking order across thicknesses. The two neutral trilayer polytypes, which remain indistinguishable by surface potential measurements and second harmonic generation, exhibit markedly different UPL intensities. This sensitivity persists in thicker samples where multiple configurations coexist. First-principles calculations suggest that the intensity contrast originates primarily from the layer confinement of the annihilating excitons, while energy matching to the $Γ$ final-state manifold provides additional intensity selectivity. Power-dependent spectroscopy further disentangles two distinct annihilation channels originating from different dark exciton valleys, identified through their contrasting intensity scaling and opposite density-induced energy shifts. Crucially, the annihilation process doubles the energy separation of nearly degenerate dark excitons while converting their weak emission into bright signal, providing experimental access to valley-specific responses that are obscured in direct dark-exciton spectroscopy. Our findings demonstrate that ferroelectric configurations provide a new degree of freedom for controlling nonlinear optical processes.

cond-mat.mtrl-sci↗

arXiv2026

Logographic Character Visual Pretraining via Semantic-based Contrastive Learning

Current deep learning-based character vision studies, e.g., text recognition, character image denoising, and historical text completion, are offering new solutions for learning, managing, and utilizing character resources. However, the performance of these studies peaks only with large and balanced datasets, which is a rarity with real-world character datasets, especially for logographic character languages, e.g., Chinese. The imbalance in data distribution of logographic characters is a common issue due to differences in character usage frequency and new characters being continuously created. In this paper, we propose a novel method for logographic character recognition, which introduces a multi-modal learning approach using visual semantics and contextual semantics of characters. A novel pre-training strategy is designed to enhance deep visual representations, especially for datasets suffering from issues of imbalanced and rare instances, by extracting the contextual semantics of each character from the corresponding language models. We conduct experiments across various datasets to evaluate our character recognition method and further validate the contrastive pre-training strategy by several downstream tasks. Experimental results demonstrate the superiority of our method compared to state-of-the-art methods.

cs.CV↗

arXiv2026

Active View Selection with Perturbed Gaussian Ensemble for Tomographic Reconstruction

Sparse-view computed tomography (CT) is critical for reducing radiation exposure to patients. Recent advances in radiative 3D Gaussian Splatting (3DGS) have enabled fast and accurate sparse-view CT reconstruction. Despite these algorithmic advancements, practical reconstruction fidelity remains fundamentally bounded by the quality of the captured data, raising the crucial yet underexplored problem of X-ray active view selection. Existing active view selection methods are primarily designed for natural-light scenes and fail to capture the unique geometric ambiguities and physical attenuation properties inherent in X-ray imaging. In this paper, we present Perturbed Gaussian Ensemble, an active view selection framework that integrates uncertainty modeling with sequential decision-making, tailored for X-ray Gaussian Splatting. Specifically, we identify low-density Gaussian primitives that are likely to be uncertain and apply stochastic density scaling to construct an ensemble of plausible Gaussian density fields. For each candidate projection, we measure the structural variance of the ensemble predictions and select the one with the highest variance as the next best view. Extensive experimental results on arbitrary-trajectory CT benchmarks demonstrate that our density-guided perturbation strategy effectively eliminates geometric artifacts and consistently outperforms existing baselines in progressive tomographic reconstruction under unified view selection protocols.

cs.CV↗

arXiv2026

Learning light scattering from operator parameter spaces to Galerkin-consistent solution spaces

Efficient and generalizable full-wave simulation is essential for nanophotonic analysis and inverse design, yet existing methods face a tradeoff between the high computational cost of numerical solvers and the limited generalizability of neural operator models for complex optical scattering. Here, we introduce FEMONet, a finite-element-constrained operator-learning framework that learns light scattering from an operator parameter space to a Galerkin-consistent solution space. The operator parameter space encodes the physical entities defining a wave-equation problem, while the variational weak form links this space to the coordinate and physical solution spaces. Integrated with operator-learning networks, FEMONet extends classical solvers from isolated problem instances to parameterized scattering operators. To our knowledge, FEMONet represents the first Galerkin-consistent operator-learning framework for complex-valued optical scattering, grounded in the variational weak form of the governing vector wave equations. Finite-element discretization absorbs spatial derivatives into assembled stiffness matrices and load vectors, removing coordinate-based derivatives of the neural-network output from the physics loss and improving training efficiency. By predicting finite-element expansion coefficients rather than unconstrained field values, the Galerkin-consistent formulation preserves compatible trial and test spaces, achieving high accuracy, stable training, and generalization across dielectric, metallic, arrayed, plasmonic, and three-dimensional nanophotonic structures.

physics.optics↗

arXiv2026

Learning Thermoelectric Transport from Crystal Structures via Multiscale Graph Neural Network

Graph neural networks (GNNs) are designed to extract latent patterns from graph-structured data, making them particularly well suited for crystal representation learning. Here, we propose a GNN model tailored for estimating electronic transport coefficients in inorganic thermoelectric crystals. The model encodes crystal structures and physicochemical properties in a multiscale manner, encompassing global, atomic, bond, and angular levels. It achieves state-of-the-art performance on benchmark datasets with remarkable extrapolative capability. By combining the proposed GNN with \textit{ab initio} calculations, we successfully identify compounds exhibiting outstanding electronic transport properties and further perform interpretability analyses from both global and atomic perspectives, tracing the origins of their distinct transport behaviors. Interestingly, the decision process of the model naturally reveals underlying physical patterns, offering new insights into computer-assisted materials design.

cond-mat.mtrl-sci↗

arXiv2026

FreeOrbit4D: Training-Free Arbitrary Camera Redirection for Monocular Videos via Foreground-Complete 4D Reconstruction

Camera redirection aims to replay a dynamic scene from a single monocular video under a user-specified camera trajectory. However, large-angle redirection is inherently ill-posed: a monocular video captures only a narrow spatio-temporal view of a dynamic 3D scene, providing severely limited observations of the underlying 4D world. The key challenge is therefore to recover a complete and coherent representation from this limited input, with consistent geometry and motion. While recent diffusion-based methods achieve impressive visual generation quality, they often break down under large-angle viewpoint changes far from the original trajectory, where missing visual grounding leads to severe geometric ambiguity and temporal inconsistency. We present FreeOrbit4D, an effective training-free framework that tackles this ambiguity by recovering a foreground-complete 4D proxy as structural grounding for video generation. We obtain this proxy by decoupling foreground and background reconstructions: we unproject the monocular video into a static background and partial foreground point clouds in a unified global space, then use an object-centric multi-view diffusion model to synthesize multi-view images and reconstruct complete foreground point clouds in canonical object space. By aligning the canonical foreground point cloud to the global scene space via dense pixel-synchronized 3D-3D correspondences and projecting the foreground-complete 4D proxy onto target camera viewpoints, we provide geometric scaffolds that guide a conditional video diffusion model. Extensive experiments show that FreeOrbit4D produces more faithful and temporally coherent redirected videos under challenging large-angle trajectories, and our proxy further enables applications such as edit propagation and 4D data generation. Project page: https://freeorbit4d.vision.ischool.illinois.edu/

cs.CV↗

arXiv2026

Inducing Overthink: Hierarchical Genetic Algorithm-based DoS Attack on Black-Box Large Language Reasoning Models

Large Reasoning Models (LRMs) are increasingly integrated into systems requiring reliable multi-step inference, yet this growing dependence exposes new vulnerabilities related to computational availability. In particular, LRMs exhibit a tendency to "overthink", producing excessively long and redundant reasoning traces, when confronted with incomplete or logically inconsistent inputs. This behavior significantly increases inference latency and energy consumption, forming a potential vector for denial-of-service (DoS) style resource exhaustion. In this work, we investigate this attack surface and propose an automated black-box framework that induces overthinking in LRMs by systematically perturbing the logical structure of input problems. Our method employs a hierarchical genetic algorithm (HGA) operating on structured problem decompositions, and optimizes a composite fitness function designed to maximize both response length and reflective overthinking markers. Across four state-of-the-art reasoning models, the proposed method substantially amplifies output length, achieving up to a 26.1x increase on the MATH benchmark and consistently outperforming benign and manually crafted missing-premise baselines. We further demonstrate strong transferability, showing that adversarial inputs evolved using a small proxy model retain high effectiveness against large commercial LRMs. These findings highlight overthinking as a shared and exploitable vulnerability in modern reasoning systems, underscoring the need for more robust defenses.

cs.CR↗

arXiv2026

Zero-Shot Chinese Character Recognition via Global-Local Dual-Branch Alignment and Hierarchical Inference

Chinese character categories are extremely large, and unseen characters frequently arise in open-world scenarios, making zero-shot Chinese character recognition an important yet challenging problem. Existing IDS-based retrieval methods usually encode a character image and its ideographic description sequence into a single global vector for matching. Although efficient, such holistic alignment often under-models local component differences. Moreover, directly introducing patch-token level fine-grained interaction suffers from both the noise of structural operators in IDS and the high cost of full-candidate retrieval.To address these issues, we propose a Global-Local Hierarchical Perception Network (GL-HPN), which jointly learns global and local representations of character images and IDS sequences within a unified cross-modal alignment framework. The global branch supports efficient coarse recall, while the local branch improves component-level discrimination through patch-token interaction. We further introduce a structure filtering mask to suppress structurally meaningful but visually non-entity IDS operators in local similarity aggregation. On top of this, we design a coarse-to-fine hierarchical inference strategy that performs global retrieval over the full candidate set and local reranking only on Top-$K$ candidates, followed by parameter-free multiplicative fusion of normalized posterior scores. Experimental results show that GL-HPN achieves competitive performance across multiple zero-shot splits, performs especially well under low-resource settings, and substantially reduces the inference cost of large-scale candidate retrieval.

cs.CV↗

arXiv2026

Salca: A Sparsity-Aware Hardware Accelerator for Efficient Long-Context Attention Decoding

Long contexts improve capabilities of large language models but pose serious hardware challenges: compute and memory footprints grow linearly with sequence length. Particularly, the decoding phase continuously accesses massive KV cache, dramatically increasing bandwidth and computing pressure. Existing accelerators are primarily designed and evaluated for short contexts. They suffer from significant performance degradation when processing long contexts. To bridge this gap, we identify the major bottleneck and present a hardware accelerator for long context attention decoding via hardware-software co-design. On the software side, we propose dual-compression dynamic sparse attention. It combines ultra-low-precision quantization with feature sparsity to minimize prediction overhead. A hardware-friendly approximate Top-K selection further reduces filter complexity from $O(n \log k)$ to $O(n)$. On the hardware side, we deeply optimize compute and memory access to tackle bottlenecks from intricate interplay between sparse attention and long contexts, and establish a performance model to derive the optimal co-design scheme. The resulting hardware adopts a fully pipelined parallel architecture and achieves $O(n)$ efficiency even for long sequences. Experiments show that our design delivers $3.82\times$ speedup and $74.19\times$ energy efficiency over A100. Compared to SOTA accelerators, this is the first ASIC accelerator that efficiently supports long context inference, with at least $3.5\times$ higher throughput and $2.08\times$ better energy efficiency.

cs.AR↗