arXiv subjects

Wei Xu

Publications and source records attributed to Wei Xu.

At least 19 recordsLinked to original sources

arXiv2026

Distillation as Probability Transport: Routed On-Policy Distillation

On-policy distillation (OPD) transfers teacher knowledge on student-generated trajectories, but efficient sampled objectives reduce the teacher distribution to scalar credit on individual tokens. Such credit indicates whether a token should gain or lose probability, yet leaves the corresponding redistribution unspecified. We recast OPD as teacher-guided probability transport and propose RouteOPD (Routed On-Policy Distillation), which decomposes local teacher--student disagreement into student-excess sources and teacher-deficit destinations and couples them into explicit transport pairs. RouteOPD optimizes pairwise log-odds toward jointly realizable targets obtained from a bounded teacher potential, while adapting the transport budget to the concentration of teacher demand. This formulation directs updates toward teacher-preferred destinations and controls their magnitude within a single transport operator. Experiments across four teacher--student settings and four mathematical-reasoning benchmarks demonstrate that RouteOPD consistently outperforms sampled reverse-KL OPD, with improvements accompanied by higher routing fidelity and lower background leakage. These results demonstrate the effectiveness of explicitly modeling probability transport in on-policy distillation.

cs.LG↗

arXiv2026

HeadWiseKV: Budgeted Per-Head Cache Residency for Hybrid Long-Context Language Models

Long-context inference retains a growing key--value (KV) cache during decoding, which consumes substantial GPU memory and can reduce generation throughput. This bottleneck remains in hybrid language models because their residual global-attention layers can dominate context-dependent cache demand. We study how to allocate this state under an aggregate KV-residency budget. We introduce HeadWiseKV, a training-free framework that compresses the residual global KV caches of hybrid language models while preserving their native local, recurrent, and linear paths. It assigns each physical KV head a static, multilevel history window, making cache demand predictable before serving. We formulate this allocation as a restricted operational rate--distortion problem and propose SeqCalib as the core policy-generation algorithm in HeadWiseKV. SeqCalib processes layers in execution order and conditions each decision on the lower-layer policy used at deployment, thereby accounting for interactions across depth. A grouped-cache runtime materializes the selected policy as actual per-head KV residency rather than a mask over a full cache. We evaluate downstream quality across four hybrid long-context models and study physical residency and serving behavior on Qwen3.6-27B. HeadWiseKV retains near-Full-KV RULER and LoCoMo quality across the evaluated models. In the fixed-model systems study, it reduces sampled peak device memory by 8.59\% at a 112K context length and extends the largest verified successful context from 114K to 161K.

cs.AI↗

arXiv2026

Before the Script, Set the Stage: How Worldview Simulation Amplifies Psychologically Grounded Persuasion in Multi-Turn Jailbreaking

Multi-turn jailbreak attacks demonstrate that harmful intent can be distributed across dialogue, yet existing methods obscure what conversational mechanisms drive vulnerability. We introduce BLUEPRINT, a safety-evaluation framework separating a factorized social-influence strategy space from WORLDVIEWSIM, a cross-turn situational context module. Monte Carlo Tree Search optimizes turn-level combinations of 18 theory-grounded influence factors across a four-turn trajectory. Across six frontier models, BLUEPRINT achieves near-ceiling ASR on major open-weight and proprietary models, while requiring the fewest average queries (2.46). The resulting trajectories further reveal model-specific vulnerability among resistant targets: each responds to distinct influence factors and strategy transitions, yet all share a common recovery pathway-shifting toward concrete, executable task framing consistently escapes hard-refusal states. Ablations confirm operational cues matter most: making requests actionable has the largest impact, gain framing is unusually potent, and some legitimacy appeals can backfire. These findings suggest robust multi-turn safety requires monitoring not only harmful content, but also how dialogue state makes unsafe requests appear concrete and locally executable.

cs.CL↗

arXiv2026

VakyArth: Evaluating Pragmatic Competence in LLMs across Indic Languages

Real-world communication often requires pragmatic reasoning: interpreting meanings implied through context and cultural convention rather than stated literally. Existing pragmatic evaluation remains largely limited to English and high-resource languages, leaving Indic languages unexplored despite their linguistic and cultural diversity. We introduce VakyArth, the first pragmatic benchmark for Indic languages, designed as a diagnostic evaluation covering Hindi, Punjabi, Tamil, and Malayalam. VakyArth evaluates models across five phenomena: deixis, speech acts, implicature, social pragmatics, and coherence; through multiple-choice questions, natural language inference, and translation, with all items authored by native speakers. Across multilingual large language models (LLMs) of varying families and sizes, we find consistent failures on pragmatic meanings rooted in Indic linguistic and cultural conventions. Our analysis shows systematic differences across languages and tasks: MCQ accuracy exceeds NLI accuracy in all model-language combinations, translation performance does not reliably track pragmatic understanding, and Indo-Aryan languages show a translation advantage over Dravidian languages. We further show that automatic translation metrics can miss fluent but pragmatically unfaithful outputs, especially for implicature and deixis.

cs.CL↗

arXiv2026

Synthetic Worlds for Temporal Evaluation and Knowledge Updating in LLMs

Large language models (LLMs) rely on static pretraining corpora, causing their knowledge to become outdated over time. Existing approaches for evaluating knowledge edits either suffer from rapid contamination or rely on counterfactual edits that conflict with rigid existing knowledge. In this work, we propose a synthetic, simulation-driven framework for studying knowledge insertion in LLMs. We introduce {\sc ParallelEvents}, a benchmark of fictional yet realistic future worlds that generates coherent event trajectories for controlled evaluation, avoiding contamination while preserving consistency. Building on this dataset, we develop {\sc Synapse}, a training framework that uses model-generated data to update model parameters via mid-training and instruction tuning. This synthetic pipeline enables scalable knowledge integration without costly human-curated data. Empirically, {\sc Synapse} outperforms existing methods by 14.23\%, demonstrating that simulation-based synthetic training leads to robust and coherent knowledge insertions.

cs.CL↗

arXiv2026

CoCoA: Context-Conditional Cultural Alignment for Large Language Models

Large Language Models (LLMs) often favor Western-associated entities across cultural contexts. Conventional debiasing methods aim for uniform neutrality, but cultural bias mitigation demands context-conditional behavior, preferring culturally appropriate entities when cultural cues are present and remaining neutral when they are absent. We propose CoCoA (Context-Conditional Cultural Alignment), a framework that learns this behavior through dual-context training on the same entity pairs under contexts with and without cultural cues. CoCoA combines a contrastive alignment objective with calibration and drift regularization, optimized through goal-aware gradient reconciliation. We evaluate CoCoA on CAMeL and Camellia, two entity-centric cultural bias benchmarks, across ten language settings and four LLMs. CoCoA reduces the Cultural Bias Score from 43 to 24 on average while maintaining near-neutral preferences at 50.2, with minimal impact on general performance across five standard benchmarks. These findings highlight that effective cultural alignment requires context-conditional modeling rather than uniform debiasing, and establish a new direction for mitigating entity-centric cultural bias in LLMs.

cs.CL↗

arXiv2026

Global Existence of classical solutions to 3D nonlinear Klein-Gordon equations with low-regularity initial data

This paper studies global existence for the Cauchy problem of nonlinear Klein-Gordon equations in three space dimensions, strictly within the regularity regime of classical local existence. We prove it via higher-order and lower-order energy estimates. The proof relies on two key ingredients. The first is due to the work of Georgiev and Popivanov, which reduces quadratic nonlinearities to cubic terms plus ghost-energy-controllable terms, securing lower-order estimates. The second is a sharp Klainerman-Sobolev-type inequality without the scaling operator, established herein, which yields enough time decay for derivatives up to second order; integration by parts then controls derivative loss terms in the higher-order estimates.

math.AP↗

arXiv2026

Should I Use This Synthetic Dataset for Training? How to Test with Minimal Real Data

Digital twins (DTs) and learned world models are increasingly used to generate synthetic data that augment the scarce real datasets available for training artificial intelligence (AI) models in engineering systems. Owing to the inevitable simulation-to-reality (sim-to-real) gap, however, augmentation may fail to improve the performance of the trained model on the real data distribution. This paper addresses the resulting decision problem: Given a real dataset, a candidate synthetic dataset, and a fixed learning algorithm, decide whether training on the augmented dataset improves the true, population-level performance, while consuming as few real test data points as possible. Two formulations are considered: a direct test on the mean loss difference between the two trained models, and a symmetry-based test on the paired loss difference, which trades a stronger null assumption for faster evidence accumulation. For the latter, we introduce the {adaptive e-process sign-flip test} (aeSFT), a doubly adaptive procedure that adapts both the number of Monte Carlo sign-flip rounds, and hence the computational cost, and the amount of real test data consumed. aeSFT yields anytime-valid Type-I error control, with no need to pre-specify the test-set size. Experiments on a synthetic-data classification task, a DT-aided wireless packet-scheduling task, and a radio-map prediction task show that aeSFT identifies useful synthetic data using substantially fewer real test samples than mean-based sequential testing, matches the power of fixed-sample sign-flip testing and the paired $t$-test, while keeping the false-positive rate below the target level.

cs.AI↗

arXiv2026

Manifold4D: Denoising on Point Cloud Rendered Manifolds for Video Re-shooting

Video re-shooting re-renders a monocular video of a dynamic scene along a user-specified camera trajectory, and the dominant recipe supplies the target geometry explicitly: per-frame depth lifts the source video into a 4D point cloud, which is rasterized along the trajectory into a point cloud render. Because the render and the source video are both handed to the network as visual conditions, they compete at every denoising step, leaving the model with a trust dilemma --- how much of the render to believe --- which can degrade trajectory control or visual quality on data outside the training distribution. We argue that a render already pixel-aligned with the target view does not need to be supplied as an explicit conditioning stream at all. We propose MANIFOLD4D, which injects the render directly into the initial noise of flow matching, so that generation no longer departs from standard Gaussian noise but from a new noise manifold carrying geometric information, leaving the source video as the only visual condition. The render is thus used exactly once, and the network is never asked to learn how to read it; in subsequent denoising steps the model can focus on the source video. On our DAVIS-Traj benchmark and on the Vista4D evaluation set, MANIFOLD4D attains the best camera-control accuracy on every metric, lowering rotation error by 25% and 27% and translation error by up to 32% over the strongest baseline, while matching it in video fidelity and leading on real-world novel-view photometric quality. In a user study, our method achieves clear advantages in trajectory following and dynamic consistency. The gap widens as the yaw amplitude grows past the training range, and the model still recovers correct dynamic motion from the source video when the render is deliberately corrupted, confirming that the geometric prior guides generation without overriding it.

cs.CV↗

arXiv2026

Scaling phoneme-based TTS augmentation for ASR: A unified pipeline and controlled study

Synthetic speech offers scalable supervision for automatic speech recognition (ASR), but its benefit depends on text selection, reference speech, and augmentation scale. We present a phoneme-based TTS-to-ASR pipeline using a single TTS model jointly trained from scratch on Arabic, French, Italian, and Portuguese with the F5-TTS architecture and language-monolingual ASR systems cover 13 test sets. Across the synthesis-scale sweep, random augmentation improves over matched real-only continuation on 11 sets. In the selection comparison, PFGS improves over real-only training on 12 sets and over random selection on nine, with a maximum relative WER reduction of 19.3% against random selection. With target texts and synthesis counts fixed, reference-speech filtering reduces absolute WER by 0.29 and 0.59 points on Italian and French Common Voice, respectively. These findings support treating TTS augmentation as a synthetic-corpus construction problem, rather than merely a question of generation scale.

cs.CL↗

arXiv2026

Probing three-dimensional structures of complex colloidal quantum dots at the single-atomic level

Colloidal quantum dots (QDs) are promising optoelectronic materials due to their size-tunable properties, yet their three-dimensional (3D) quantum confinement makes electronic states highly sensitive to structural and chemical heterogeneity, which critically impacts their optoelectronic performance. Accurately resolving the 3D atomic structure with sub-angstrom precision is thus essential for rational design. Here, we applied atomic electron tomography (AET) to determine, for the first time, the 3D atomic structure of complex core/shell QDs, resolving over 14,000 atoms per particle. Our reconstructions reveal surface morphology, eccentric cores, and nearly atomically abrupt heterovalent interfaces and identify anisotropic shell growth directed by twin boundaries. Utilizing an AET-derived atomic structure, we performed large-scale quantum mechanical calculations to uncover an orientation-dependent strain accommodation mechanism where the heterogeneous strain is compensated at interfaces and twin boundaries. Furthermore, our results reveal strain-induced localized states near the band edge, which contribute to the key features of the experimental ensemble absorption spectrum. This work sets a new benchmark for atomic-level characterization, establishing a powerful framework for the rational design of next-generation nanomaterials.

cond-mat.mtrl-sci↗

arXiv2026

Uniform Local Asymptotics for L\'evy Processes with Subexponential Jumps

This paper is devoted to unifying the uniform local large-deviation asymptotics for a centered L\'evy process $X$ with subexponential jumps. Our results assert that for any $\theta,\delta_0>0$ and $K\geq0$, $$\lim_{t\to\infty}\sup_{x\geq\theta t}\sup_{|y|\leq Kb(x)}\sup_{\delta\in[\delta_0,\infty]}\sup_{0<s\leq t}\bigg|\frac{\mathbf P\big(X_s\in(x-y,x-y+\delta]\big)}{s\cdot\mathbf P\big(X_1\in(x,x+\delta]\big)}-1\bigg|=0,$$ where the natural-scale function $b$ satisfies a polynomial growth condition. This provides a continuous-time and simultaneously uniform analogue of the results of Denisov et al. [Ann. Probab., 2008], while being established under a weaker moment assumption.

math.PR↗

arXiv2026

Federated Unlearning Over Wireless Networks

To comply with stringent data privacy regulations, federated unlearning (FU) has emerged as a critical paradigm. However, its implementation over wireless networks introduces severe communication latency and reliability challenges due to iterative calibration requirements and physical-layer channel uncertainties. In this paper, we investigate the problem of delay minimization for federated unlearning networks (FUN). Specifically, we establish a comprehensive system model that jointly incorporates the convergence behavior of the FUN algorithm, local device computation dynamics, and a worst-case robust transmission model operating under bounded channel state information (CSI) error. To solve the resulting non-convex joint resource allocation problem, we propose an efficient iterative algorithm. By exploiting the monotonicity and convexity properties of the system constraints, the problem is decomposed via a uniform scan over the local accuracy parameter, within which the optimal delay, bandwidth, power, and computation frequency are determined utilizing nested bisection and golden-section searches. Both theoretical analysis and extensive numerical results demonstrate that the proposed algorithm achieves polynomial complexity and significantly reduces the overall unlearning completion time compared to conventional baseline schemes.

cs.IT↗

arXiv2026

Transfer Learning-Enabled Distortion Compensation for Amplitude-Phase-Time Block Modulation-Based Nonlinear Single-Carrier Wireless Communications

Power amplifier (PA) nonlinearity and memory effects significantly limit the spectral compliance, reliability, and energy efficiency of communication systems. To address this, we propose a transfer-learning-enabled, fully digital transceiver-cooperative method for amplitude-phase-time block modulation (APTBM)-based nonlinear single-carrier transmission under adjacent channel leakage ratio (ACLR) constraints. At the transmitter, iterative clipping and filtering (ICAF) and static digital pre-distortion (SDPD) act jointly to reduce signal peaks and suppress spectral regrowth without requiring wideband feedback. At the receiver, the inherent amplitude-phase constraints of APTBM provide weakly supervised prior knowledge for offline inverse-model pretraining, which is followed by the online few-shot adaptation of a lightweight digital post-distortion (DPoD) network. Subsequently, a cascaded DPoD and clipping-noise cancellation scheme systematically compensates for residual distortions induced by both the PA and ICAF. Simulation and measurement results demonstrate reliable transmission at an input back-off of approximately 2 dB under a 30-dBc ACLR constraint. Furthermore, the proposed DPoD approach significantly reduces online training time and computational overhead, delivering a performance gain of over 2 dB compared to conventional learning-based DPoD schemes.

eess.SP↗

arXiv2026

XGait: A Multi-Modality Wireless Sensing Dataset for Indoor Human Tracking and Identification

Wireless sensing has emerged as a promising approach for tracking and identification using commodity Internet of Things devices. However, the features derived from a single wireless modality are often fragile to variations in environmental layouts and walking trajectories. Furthermore, most existing studies are based on datasets collected in specific scenarios with limited trajectory diversity and sensing modalities, preventing a robust evaluation of system generalization. \textcolor{blue}{To address this gap, we introduce \textbf{XGait}, a multi-modality wireless sensing dataset that synchronously captures human walking using Wi-Fi and acoustic transceivers across three indoor scenarios, with vision-based measurements serving as ground truth. Specifically, XGait contains more than 22K walking samples from 27 participants, covering diverse directions and trajectories to support both indoor tracking and identity recognition. To bridge the heterogeneity of wireless sensing modalities, we propose a unified Doppler spectrogram representation that maps Wi-Fi and acoustic signals into a shared time--frequency space, along with a standardized benchmark pipeline for pre-processing, temporal alignment, and feature construction, enabling reproducible evaluation and systematic cross-modal analysis. Extensive evaluations demonstrate that Wi-Fi and acoustic sensing exhibit complementary strengths, particularly under complex trajectories and challenging propagation conditions, thereby paving the way for novel research in the field of multi-modality wireless sensing.} The dataset and code are available at https://github.com/warrior-087/XGait.

cs.HC↗

arXiv2026

What Language Does and What the Evidence Supports: A Functional Role Taxonomy and Evidence Audit of Language Grounding in Embodied Agents

Foundation models place language throughout embodied agents, but its presence does not show what it contributes or how well that contribution is grounded. This survey separates these two questions. We define five non-exclusive functional roles for language: Specification, Embodied Representation, Action Orchestration, Grounding Regulation, and Execution Coupling. For each role, we trace the path from linguistic content to its embodied consumer and identify the observations or interventions that can test the claimed responsibility. Applying this framework to the reviewed literature reveals a recurring gap between functional use and evidential support. Interpretable or revised linguistic intermediates may be incorrect, go unused, or fail to affect later behavior. Even when actions are directly conditioned on language, system-level success does not by itself isolate language's contribution. We therefore evaluate grounding claim by claim, asking whether the reported evidence supports the specific responsibility assigned to language. Using role claims rather than architectures as the unit of comparison allows us to compare modular and end-to-end embodied agents without extending conclusions beyond the reported evidence.

cs.CL↗

arXiv2026

SiMDex: Mining Similar Egocentric Videos for Cross-Embodiment Dexterous Manipulation

Recent years have witnessed an explosive trend of scaling ego-centric human videos for robot manipulation, yet it remains unclear which data actually benefits dexterous manipulation. We present SiMDex, a similarity-based data mining framework that casts human data selection for VLA post-training in dexterous manipulation as a recommendation problem. For each robot demonstration, SiMDex employs a three-layer recall-ranking-re-ranking pipeline to extract task-relevant subsets from a pool of ~32M egocentric human samples, operating in a morphology-agnostic action space that requires no changes to VLA architecture or training. Against a strong baseline trained with an equal amount of randomly sampled human data, SiMDex uses only ~1.49M mined samples (<5% of the pool) yet improves the overall success rate from 47.7% to 61.1%, showing that selective curation outperforms indiscriminate data mixing.

cs.RO↗

arXiv2026

Diarization-Guided Qwen-ASR Adaptation for Multilingual Two-Speaker Conversational Speech

This paper describes our self-designed system for Task 1 of the MLC-SLM 2026 Challenge for multilingual two-speaker conversational speech. The system combines a modular speaker diarization front end with a challenge-adapted Qwen3-ASR-1.7B recognizer. The diarization front end performs voice activity detection, subsegment generation, CAMPPlus speaker embedding extraction, two-speaker spectral clustering, and RTTM-based audio segmentation. The resulting speaker-attributed segments are grouped by language or region and decoded by the adapted ASR model. For ASR adaptation, we first perform supervised full fine-tuning on the official training data, then apply LoRA fine-tuning with synthetic speech generated by a three-pipeline TTS-based synthetic speech augmentation framework, and finally refine the model using GRPO reinforcement learning with rewards based on WER/CER and penalties for hallucination, repetition, and length deviation. On the official development set, the full system achieves an average tcpMER of 23.70, reducing the error rate by 6.83 absolute points relative to the released Qwen-ASR-1.7B performance. On the final evaluation set, the system achieves an average tcpMER of 17.97. Ablation results show that supervised fine-tuning provides the largest gain, while synthetic-speech LoRA adaptation and reinforcement learning further improve robustness.

cs.CL↗