arXiv subjects

Wei Liu

Publications and source records attributed to Wei Liu.

At least 109 records · Page 6Linked to original sources

arXiv2026

DanceOPD: On-Policy Generative Field Distillation

Modern image generation demands a single model that unifies diverse capabilities, including text-to-image (T2I), local editing, and global editing. However, these capabilities are rarely naturally aligned and often conflict. For instance, editing tends to degrade T2I performance, while global and local editing interfere with each other. Consequently, effectively composing these capabilities has become a central challenge for image generation model training. To tackle this, we introduce DanceOPD, an on-policy generative field distillation framework for flow-matching models that routes each sample to one capability field, queries one low-noise student-induced state, and trains with a simple velocity MSE objective. With each capability source defined as a velocity field over the shared flow state space, the student learns from fields queried on its own rollout states to compose expert capabilities. This formulation also absorbs operator-defined fields such as classifier-free guidance. Comprehensive experiments on T2I, editing, realism-field absorption, and CFG absorption show that our approach improves multi-capability composition, strengthening target capabilities while preserving anchor generation quality. We believe this work establishes a practical route for generative field distillation in flow-matching models.

cs.CV↗

arXiv2026

Learning Dexterous Manipulation Using Contact Wrench Guidance From Human Demonstration

Dexterous robot manipulation can benefit from the abundance of human demonstrations, but transferring such demonstrations to robot policies remains challenging. We present Contact Wrench Guidance from Human Demonstration in Robotic Dexterous Manipulation (CHORD), a framework for long-horizon manipulation of rigid and articulated objects with reinforcement learning. The key idea is object-centric contact wrench space guidance: we represent human and robot motions by the forces and torques they can induce on the object, enabling similarity to be measured by the induced instantaneous motions. This guidance makes reinforcement learning more scalable for contact-rich dexterous manipulation. We further introduce a large-scale simulation benchmark with 4,739 bimanual dexterous manipulation tasks, constructed from motion-capture datasets and reconstructed in-house videos. Evaluated on 1,831 benchmark tasks, CHORD achieves an average success rate of 82.12%, demonstrating strong scalability. CHORD also generalizes to whole-body manipulation from hand-only and third-person demonstrations, achieving a 90.77% success rate, and the learned policies transfer to the real world in both open-loop and closed-loop settings.

cs.RO↗

arXiv2026

CoAdapt-GUI: Joint Workflow Context and Policy Adaptation for Unseen GUI Applications

Mobile GUI agents remain brittle when deployed to applications absent from source training. We study novel-app generalization under a limited target interaction budget and without target demonstrations. We introduce CoAdapt-GUI, a test-time adaptation (TTA) framework that jointly adapts structured workflow context and policy from the agent's own target-app rollouts and rewards. The workflow context retains transferable procedures, failure modes, and verification rules while excluding app-bound source details. This separation allows reusable workflow knowledge to guide adaptation without transferring source-interface state. For policy adaptation, task-context-matched group-relative optimization updates a LoRA adapter on a frozen vision-language model. Across two unseen-app evaluations, CoAdapt-GUI reaches 45.0% on AndroidWorld-Generalization, compared with 37.5% for the reported Policy-Only TTA baseline, and raises AndroidWorld Plus performance from 38.6% to 52.9%. These results show that transfer-constrained workflow context provides substantial gains and that joint policy adaptation further improves held-out performance.

cs.AI↗

arXiv2026

Preserve More Details: Mitigating Content Drift in Real-World Image Super-Resolution

Real-world image super-resolution (Real-ISR) aims to reconstruct high-quality (HQ) images from low-quality (LQ) inputs subject to diverse real-world degradations. Recent advances have leveraged the LQ inputs and natural image priors learned by Stable Diffusion models to achieve impressive results. However, existing methods often overlook insufficient clarity of LQ inputs inevitably induce content drift in the generated HQ images. This manifests primarily as visual detail degradation and textual semantic shift, severely compromising both fidelity and perceptual quality. To address this challenge, we propose FSP-Diff, a novel one-step diffusion model featuring a dual-pathway architecture. This architecture comprises a Detail-Conditioned Pathway for injecting structured details to recover fine structures, and a Detail-Modulated Semantic Pathway that refines semantic guidance using structured details to mitigate semantic deviations. Extensive experiments on standard Real-ISR benchmarks demonstrate that FSP-Diff surpasses existing one-step diffusion methods in both quantitative and qualitative metrics.

cs.CV↗

arXiv2026

Beyond Uniform Restoration: Empowering All-in-One Restoration with Pixel-Level Multimodal Guidance

All-in-one image restoration is a unified low-level vision task that aims to effectively recover high-quality images from inputs degraded by various types and levels of corruption using a single model. Recent works have achieved remarkable progress by learning degradation-adaptive prompts or network architectures. However, these methods typically apply a uniform restoration strategy across the entire image, neglecting the fact that different regions may suffer from distinct degradation types and varying degrees of severity. In contrast, we propose to perform restoration at the pixel level, thereby enabling more fine-grained and precise control over the restoration process. Specifically, we present MGN-AIR, a novel pixel-level restoration framework for all-in-one image restoration. Our approach first learns to estimate a pixel-level visual prompt. Then, it leverages both textual and visual prompts to provide global and local degradation cues, guiding the model on where to look and how to restore at each pixel. We conduct extensive experiments on multiple all-in-one image restoration benchmarks, covering a wide range of tasks including denoising, deraining, deblurring, dehazing, desnowing, and low-light enhancement. Experimental results demonstrate that our proposed method consistently and significantly outperforms existing approaches.

cs.CV↗

arXiv2026

Learning Visual Spatial Planning from Symbolic State via Modality-Gap-Aware Self-Distillation

While Vision-Language Models excel at general multimodal understanding, they still struggle with visual spatial planning. We attribute this limitation to a perception--reasoning modality gap. Visual planning requires models to infer latent state structures from pixels and then reason over the recovered structure to produce valid actions, whereas symbolic planning directly leverages explicit representation. This discrepancy introduces two sequential bottlenecks: visual state recovery at the perception stage and multi-step planning at the reasoning stage. To address this, we propose MGSD, a two-stage modality-gap-aware self-distillation framework. First, a cold-start grounding stage establishes reliable visual state recovery before on-policy training. Second, a symbol-guided on-policy self-distillation stage transfers the privileged teacher's planning behavior to the student through token-level supervision on student-generated prefixes. Crucially, symbolic information is used only during training, while inference relies exclusively on visual inputs. Experiments on visual planning benchmarks show that MGSD consistently improves performance across different model scales, raising the macro average by 19.3% and 18.4%, respectively. The resulting models substantially reduce the gap to the upper bounds obtained with symbolic inputs. Ablation studies and diagnostic analyses further confirm that the gains arise from improvements in both visual state recovery and optimal-path reasoning. These results demonstrate that MGSD strengthens not only the recovery of actionable states from visual observations but also the ability to plan over the inferred structures. Code is available at https://github.com/Oranger-l/MGSD.

cs.AI↗

arXiv2026

CoSA: Accelerating Long-Context Inference via Proxy-Kernel Co-Designed Sparse Attention

The quadratic cost of self-attention makes long-context inference prohibitively expensive, and proxy-based block-sparse attention has become a practical remedy. Existing methods typically rely on a proxy to predict a binary sparse mask and a kernel to consume this mask and perform sparse attention computation. Such an approach is effective under moderate budgets. However, as the budget tightens, the estimated proxy inevitably drops some salient blocks, while the kernel can only apply the sparse mask mechanically, leading to an evident drop in model accuracy. We propose CoSA, a two-stage training-free Sparse Attention under proxy-kernel CO-design, which couples a Kernel-Aware Proxy (KAP) with an Ordered-Skipping Kernel (OSK). In the first stage, the KAP selects blocks under a moderate budget and produces an ordered mask that prescribes the order in which KV pages are visited in the kernel inner loop. In the second stage, the OSK applies this mask and skips more blocks under a tightened budget given online-softmax statistics. Across mainstream LLM backbones and long-context benchmarks, CoSA attains higher accuracy at lower budgets. Impressively, CoSA achieves a 4.93$\times$ attention speedup and reduces end-to-end Time-to-First-Token by 2.53$\times$ under a context length of 128K with negligible performance degradation. Code is available at https://github.com/Tencent/AngelSlim.

cs.CL↗

arXiv2026

A Mechanistic Analysis of Transformers for Dynamical Systems

Transformers are increasingly adopted for modeling and forecasting time-series, yet their internal mechanisms remain poorly understood from a dynamical systems perspective. In contrast to classical autoregressive and state-space models, which benefit from well-established theoretical foundations, Transformer architectures are typically treated as black boxes. This gap becomes particularly relevant as attention-based models are considered for general-purpose or zero-shot forecasting across diverse dynamical regimes. In this work, we do not propose a new forecasting model, but instead investigate the representational capabilities and limitations of single-layer Transformers when applied to dynamical data. Building on a dynamical systems perspective, we interpret causal self-attention as a linear, history-dependent recurrence and analyze how it processes temporal information. Through a series of linear and nonlinear case studies, we identify distinct operational regimes. For linear systems, we show that in the single-head attention-only setting, the convexity constraint imposed by softmax attention restricts the class of autoregressive operators that can be represented, leading to oversmoothing when the target dynamics require mixed-sign lag coefficients. For nonlinear systems under partial observability, attention instead acts as an adaptive delay-embedding mechanism, enabling effective state reconstruction when sufficient temporal context and latent dimensionality are available. These results help bridge empirical observations with classical dynamical systems theory, providing insight into when and why Transformers succeed or fail as models of dynamical systems.

cs.LG↗

arXiv2026

Saturation-Aware Predictive Quantization for Low-Power ECG Acquisition: A Benchmark of Taylor, Adaptive-Order, Kalman, and LSTM Predictors

Wearable electrocardiogram (ECG) monitors require energy-efficient analog-to-digital converters (ADCs), yet conventional successive-approximation-register (SAR) ADCs repeatedly resolve slowly varying most significant bits. Predictive quantization (PQ) instead estimates the next sample and quantizes only the residual, thereby reducing the required conversion depth. Its principal failure mode is residual saturation, which occurs when prediction error exceeds the residual ADC range and is irreversibly clipped. We compared four one-step-ahead predictors under a common 10-bit, saturation-aware PQ model with residual widths from 2 to 8 bits. The benchmark included first-order Taylor extrapolation, an adaptive-order predictor, a constant-velocity Kalman filter, and a two-layer long short-term memory (LSTM) network. We used an open-loop protocol in which all predictors received past original samples. This protocol isolates intrinsic prediction performance from recursive reconstruction-error propagation. Saturation rate (SR) was the primary metric, complemented by overflow energy ratio (OER), which weights each event by its squared overflow depth. On a 5,317-sample excerpt from MIT-BIH Arrhythmia Database Record 101, the Kalman predictor performed best at Br=6. It achieved 30.88 dB SNR, 2.16% SR, and 12.93% OER, compared with 28.39 dB, 2.69%, and 25.29% for Taylor extrapolation. The adaptive-order predictor achieved 29.61 dB SNR and 2.44% SR using three registers, two comparators, and no multiplier. The LSTM reached 29.28 dB SNR and did not outperform the model-based predictors on this limited-data benchmark. Under the evaluated excerpt and open-loop protocol, Br=6 provided a favorable balance between reconstruction fidelity and conversion depth. Closed-loop, multi-subject, and hardware validation are required before system-level energy or deployment claims can be made.

eess.SP↗

arXiv2026

A unified framework for equation discovery and dynamic prediction of hysteretic systems

Hysteresis is a nonlinear phenomenon with memory effects, where a system's output depends on both its current state and past states. It is prevalent in various physical and mechanical systems, such as yielding structures under seismic excitation, ferromagnetic materials, and piezoelectric actuators. Analytical models like the Bouc-Wen model are often employed but rely on idealized assumptions and careful parameter calibration, limiting their applicability to diverse or mechanism-unknown behaviors. Existing equation discovery approaches for hysteresis are often system-specific or rely on predefined model libraries, which limit their flexibility and ability to capture the hidden mechanisms. To address these challenges, this research classifies equation discovery problems for hysteretic systems and develops a unified framework in which the state-space form is reformulated, and hysteretic variables are treated as trainable parameters from data. The framework further employs symbolic regression (SR) to automatically recover explicit governing equations without relying on predefined libraries, unlike methods such as sparse identification of nonlinear dynamics (SINDy). Experimental results demonstrate that the proposed method is effective in recovering governing equations for hysteretic systems, even in a challenging Full Equation Discovery setting, where prior information is extremely limited, and solving the equations naturally enables the dynamic prediction of hysteretic systems.

cs.CE↗

arXiv2026

On the Rate of Convergence of Kolmogorov-Arnold Network Regression Estimators

Kolmogorov-Arnold Networks (KANs) approximate multivariate functions by composing univariate transformations through additive or multiplicative aggregation. We establish convergence guarantees for KANs whose univariate components are B-splines. The least-squares estimator over the KAN spline sieve attains the rate $O((\log n / n)^{2r/(2r+1)})$, uniformly over a ball of regression functions admitting a KAN representation with univariate components of Sobolev smoothness $r$; a matching lower bound of order $n^{-2r/(2r+1)}$ shows this is minimax optimal up to the logarithmic factor, which we trace to the nonlinearity of the sieve rather than to the architecture. The rate is free of the ambient dimension $d$; this dimension-free exponent reflects the assumed KAN structure of the target, not an escape from the minimax rate $n^{-2r/(2r+d)}$ on Sobolev classes over $[0,1]^d$. We derive a knot-selection rule, show that penalized selection over a dyadic knot grid attains the rate adaptively in the unknown smoothness, and show that univariate components are not identifiable under centering alone, so consistency of the fit does not imply consistency of the components. On targets of exactly known smoothness the fitted risk exponent is at least as steep as the bound in every configuration, and the predicted knot scaling and $k^{-r}$ approximation decay are checked directly.

cs.LG↗

arXiv2026

Spatial Visual Analytics for Multi-Document Summary Verification

Large language models increasingly generate summaries from collections of documents to support sensemaking and reporting, but verifying whether summary statements are grounded in source materials remains difficult. In multi-document summarization (MDS), evidence is distributed across many source documents and may be incomplete, conflicting, or missing. We present Summary Verification Space (SVS), a visual analytics system for verifying multi-document summaries through spatial document organization and coordinated provenance visualization. To support scalable verification, we investigate two alternative 2D canvas layouts: a SUMMARY-GUIDED layout that organizes documents by alignment with summary sentences, and a SOURCE-GUIDED layout that arranges documents by semantic similarity. Coordinated provenance visualization then makes relationships among summary content, source documents, and supporting evidence explicit, enabling users to trace support, contradiction, and missing evidence during verification. A task-driven usage scenario illustrates an auditing workflow in which users use the layouts to locate relevant documents and coverage gaps, then inspect linked claims and source evidence to make their own grounding judgments. In a comparative study with provenance held constant, both spatial layouts improved aggregate accuracy and reduced workload relative to a linear baseline, with the clearest gains on relevance tasks. The SUMMARY-GUIDED layout provides the strongest overall balance of accuracy, efficiency, confidence, and workload.

cs.HC↗

arXiv2026

UNIFUSION: Adapting Autoregressive Language Models into Discrete Diffusion under a Unified Reverse-Rate Objective

Existing methods mainly adapt pretrained autoregressive (AR) language models to masked diffusion, whereas we directly adapt them to uniform-noise diffusion, where every token remains editable during sampling. However, adapting AR checkpoints across corruption kernels remains challenging because existing DLMs use different objectives and prediction parameterizations. We establish connections among SEDD, MDLM/GIDD, M2S, and Neural CTMC by expressing their conditional losses as a single generalized Kullback--Leibler objective over model reverse rates. We further derive conversions from clean-token predictions to concrete-score, posterior-mean, and exit-rate/jump parameterizations, yielding a shared $x_0$ interface that supports switching between mask and uniform kernels. Building on these connections, we propose \ours{}, a simple continual pre-training approach for directly adapting pretrained GPT2 checkpoints to uniform-noise diffusion. Through systematic evaluation of 124M- and 355M-parameter models, we show that \ours{} steadily improves the trade-off between generative perplexity (GenPPL) and unigram entropy as the sampling budget increases from 16 to 256 steps. At 256 steps, \ours{}-S and \ours{}-M achieve GenPPL/entropy pairs of $97.783/5.2626$ and $71.516/5.6669$, respectively; no evaluated model at the same scale simultaneously outperforms \ours{} on both metrics. At both scales, \ours{} also achieves the highest WinoGrande, SIQA, and BBH accuracy among the compared diffusion models.

cs.LG↗

arXiv2026

RM-Distiller: Exploiting Generative LLM for Reward Model Distillation

Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. Due to the difficulty of obtaining high-quality human preference annotations, distilling preferences from generative LLMs has emerged as a standard practice. However, existing approaches predominantly treat teacher models as simple binary annotators, failing to fully exploit the rich knowledge and capabilities for RM distillation. To address this, we propose RM-Distiller, a framework designed to systematically exploit the multifaceted capabilities of teacher LLMs: (1) Refinement capability, which synthesizes highly correlated response pairs to create fine-grained and contrastive signals. (2) Scoring capability, which guides the RM in capturing precise preference strength via a margin-aware optimization objective. (3) Generation capability, which incorporates the teacher's generative distribution to regularize the RM to preserve its fundamental linguistic knowledge. Extensive experiments demonstrate that RM-Distiller significantly outperforms traditional distillation methods both on RM benchmarks and reinforcement learning-based alignment, proving that exploiting multifaceted teacher capabilities is critical for effective reward modeling. To the best of our knowledge, this is the first systematic research on RM distillation from generative LLMs.

cs.CL↗

arXiv2026

An on-chip programmable mechano-quantum transducer

Solid-state spin defects encode local perturbations as measurable shifts in spin-transition frequencies, but mechanical actuation and quantum readout remain physically separated, resulting in a discrete measurement setup. Integrating these functions requires an on-site mechano-quantum interface that programs the lattice state of a defect host and quantitatively maps it onto the spin Hamiltonian. Here we first report an on-chip programmable mechano-quantum transducer (OCPMQT) that integrates voltage-defined micromechanical actuation with in situ spin-frequency readout in a two-dimensional van der Waals quantum-defect host. Mechanically programmed lattice states are encoded as shifts in the axial zero-field splitting parameter and resolved by optically detected magnetic resonance (ODMR) spectroscopy. Within a chip volume of 2.05*10^-2 cm^3, the transducer accesses ODMR-inferred strains as low as 0.0080% and delivers a volumetric force density of approximately 2.6*10^4 N*m^-3. A micromechanical-to-spin-Hamiltonian framework links on-chip electromechanics, interfacial strain transfer, and strain-spin coupling, enabling the electrical control micromechanical input to be measured directly as spin-frequency response.

cond-mat.mes-hall↗

arXiv2026

Response Morphologies of a Canonical Fluctuation Diagnostic Across Ehrenfest Phase Transitions

Microcanonical inflection-point analysis identifies MIPA-type higher-order transition structures from derivatives of the microcanonical entropy. Until recently, however, there was no canonical formulation for probing the corresponding fluctuation-level behavior directly from measurable energy fluctuations without reconstructing the density of states. Previous work addressed this limitation by introducing a canonical fluctuation diagnostic that quantifies energy-fluctuation asymmetry. Here, we investigate how this diagnostic behaves across representative first-, second-, and third-order phase transitions in the Ehrenfest classification. We consider the eight-state Potts model, the two-dimensional Ising model, and ideal three-dimensional Bose--Einstein condensation. The Potts and Ising systems exhibit a robust paired-extremum structure of the diagnostic near their respective transition regions, whereas ideal Bose--Einstein condensation exhibits a discontinuous response at the condensation temperature in the thermodynamic limit. These results show that a single fluctuation-based observable can display qualitatively distinct response morphologies across thermodynamic regimes, reflecting sensitivity to the nature of the underlying phase transition rather than serving as a direct classifier of Ehrenfest order.

cond-mat.stat-mech↗

arXiv2026

Hidden Failure Modes of Gradient Modification under Adam in Continual Learning, and Adaptive Decoupled Moment Routing as a Repair

Many continual-learning methods modify gradients upstream (e.g., projection, penalty rescaling, replay mixing) while treating Adam as a neutral backend. We show this composition has a hidden failure mode. In a high-overlap, non-adaptive 8-domain continual LM, all shared-routing projection baselines collapse close to vanilla forgetting (12.5--12.8 vs. 13.2). A 0.5% replay buffer is the strongest shared alternative but still reaches 11.6, while fixed-strength decoupling falls below vanilla at 14.1. Only adaptive decoupled routing remains stable at 9.4, improving over vanilla by 3.8 units. On a 16-domain stream, its gain over the strongest shared-routing projection baseline grows to 4.5--4.8 units. The failure is largely invisible on clean benchmarks. We explain this effect through Adam's second-moment pathway: in the tested regime, projection induces a 1/(1-alpha) inflation of the old-direction effective learning rate, matching measurements within 8% across eight alpha values. The same conflict appears with penalty methods, replay mixing, and at 7B scale under LoRA. Our fix routes the modified gradient only to the first moment while preserving magnitude-faithful second-moment statistics, with overlap-aware adaptive strength. This simple change is the only tested configuration that consistently avoids collapse across methods, optimizers, and scale.

cs.LG↗

arXiv2026

AuditRepairBench: A Paired-Execution Trace Corpus for Evaluator-Channel Ranking Instability in Agent Repair

Agent-repair leaderboards reorder under evaluator reconfiguration, and a measurable share of the reordering is produced by methods that consult evaluator-derived signal during internal selection of candidate repairs. We document this failure mode on a public leaderboard and release AuditRepairBench, a paired-execution trace corpus of 576,000 registered cells (96,000 executed) that operationalizes evaluator-channel-blocking ranking instability within a declared observability boundary. A modular screening architecture decides pathway-blocking through four interchangeable implementations, a learned influence proxy, a rule-based channel-exposure ratio that uses no trained model, a counterfactual sensitivity proxy, and a sparse human-audit proxy, combined into a screening posterior that feeds a cell-level flip functional, a set-valued label, a stratified system score, and a set-valued leaderboard. The resource is supported by mechanism-anchored validation on an 80-case source-level channel-surgery subset, an independent-discovery protocol under which two annotator groups separated from the pipeline developers discover coupling patterns blinded to the screening design and the frozen ensemble attains pooled AUROC 0.83 on their 79 cases, implementation robustness, uncertainty propagation that raises 95% coverage from 0.81 to 0.95, and forward transfer with pooled community-evaluator Spearman \r{ho} = 0.65. Screening-guided blinding patches reduce rank displacement by 55--74% (mean 62%) at fewer than 50 lines of code, whereas random channel blinding produces at most 7% reduction and generic retraining at most 13%. AuditRepairBench-Lite, a rule-only configuration on a 12,000-cell subset, preserves the leaderboard at Kendall τ = 0.88 under twenty-four GPU-hours and is the primary release artifact at 42 GB.

cs.AI↗