arXiv subjects

Xinyu Zhang

Publications and source records attributed to Xinyu Zhang.

At least 73 records · Page 4Linked to original sources

arXiv2026

Foresight Residual RL for Long-Horizon Robot Manipulation with Vision-Language-Action Models

Vision-Language-Action (VLA) policies offer strong general-purpose manipulation priors, but often fail on tight-tolerance, contact-rich assembly due to long-horizon credit assignment and subtask coupling: a state that is geometrically successful for the current skill can be brittle for downstream skills. We show this failure mode in residual reinforcement learning (RL) over a frozen VLA base policy: constant sparse success rewards improve each subtask in isolation yet yield little or no gain when skills are chained, because terminal state quality is uncontrolled. We propose Foresight Residual RL, which optimizes handoff quality by augmenting each subtask's sparse success reward with an offline-estimated foresight value -- the probability of future subtask success conditioned on the terminal state of the current subtask. Concretely, we (i) train a visual foresight predictor from images of terminal states of the base policy, labeled using downstream rollout statistics, and (ii) train residual policies via backward foresight induction, using the predictor output as a reward multiplier. On a three-phase wrench-based nut-tightening assembly task in Isaac Gym (grasp, move-insert, rotate), our method achieves 85.6% full-task success, outperforming standard subtask residual RL (54.5%) and VLA baselines, while leaving per-subtask success unchanged. These results highlight that improving long-horizon performance requires shaping which successful states are produced at each sub-task, not only whether success occurs.

cs.RO↗

arXiv2026

When Distillation Breaks Motion Control: Restoring Generative Trajectories for Fast Video Generators

Training-free motion customization imposes motion patterns from reference videos onto video generators through test-time computation. Most existing methods target full diffusion models, requiring many denoising steps and high computational cost. With the rise of efficient distilled models, a natural question arises: can test-time motion customization be applied directly to distilled generators with their accelerated sampling and efficiency gains? However, our analysis reveals that existing training-free techniques fail on distilled models. Distillation fundamentally alters the denoising dynamics that prior test-time guidance relies on, and the large denoising steps of distilled generators discard the dense intermediate states that score guidance requires, rendering existing motion control strategies incompatible with fast generation. To address this limitation, we propose MotionEcho, a novel training-free test-time distillation framework that enables motion customization for distilled video generators. The key idea is to correct the student model's sampling trajectory with restricted usage of a high-quality diffusion teacher at inference time. Teacher supervises the student's denoising by re-noising the student's endpoint onto its dense trajectory to form a motion-aligned clean endpoint, then interpolating it with the student's, while an adaptive scheduling mechanism determines when and how much teacher guidance is needed. As a result, MotionEcho restores generative trajectories for distilled video generators via lightweight, adaptive test-time teacher guidance, enabling accurate motion control without compromising generation efficiency. Extensive experiments on multiple distilled video generation models demonstrate that our method significantly improves motion fidelity and visual quality while retaining the efficiency advantages of distilled generation.

cs.CV↗

arXiv2026

From Open Loop to Closed Loop: A Test-Time Iterative Optimization Framework for Reference-Consistent Image Generation

While controllable image generation has made significant strides by incorporating visual reference conditions, existing methods predominantly operate as open-loop systems. They inject control signals in a strictly feed-forward manner, failing to guarantee strict fidelity to the reference due to the absence of active feedback and error correction mechanisms. To address this fundamental limitation, we propose a novel test-time iterative optimization framework that reformulates reference-consistent generation as a closed-loop dynamic tracking problem. By treating the pre-trained generative model as a control plant, our framework employs a sensor-controller architecture driven by a modified Proportional-Integral-Derivative (PID) algorithm. This mechanism iteratively optimizes the latent control signals at test time based on the sensed discrepancy between the generated output and the reference target. Notably, this approach is entirely training-free, model-agnostic, and integrates seamlessly around existing diffusion pipelines. Extensive evaluations across ID-preserving, pose-controlled, and depth-controlled generation tasks validate the universality of our method. Empirical results demonstrate improvements over computation-matched open-loop baselines, achieving relative performance gains of up to 25.36\% for facial similarity, alongside spatial error reductions of up to 27.71\% for pose alignment and 28.50\% for depth consistency. More broadly, this work offers a new conceptual perspective: it demonstrates that controllable generation can be effectively managed as a dynamic feedback system, bringing the rigorous principles of classical control theory into the optimization of generative models. Code is available at https://github.com/zzdrill/From-Open-Loop-to-Closed-Loop.

cs.CV↗

arXiv2026

Optimal Mixture-of-Experts Model Averaging for Conditional Generative Models

Conditional generative models have emerged as powerful tools for sampling from target conditional distributions, driving substantial advances across a wide range of scientific and applied domains. As these models proliferate, practitioners often face multiple plausible generators whose performance can vary with the task, data, or input condition. We propose an optimal model averaging framework for conditional generative models, allowing candidate generators to be combined even when they are accessible only through conditional samples without tractable densities. Specifically, we use a sample-based maximum mean discrepancy between conditional distributions, which first leads to a static model averaging method, StaticMA, assigning fixed weights to different candidates. In addition, we develop MoEMA (mixture-of-experts model averaging), an input-adaptive method that parameterizes covariate-dependent weights through a softmax neural-network gate. We establish in-sample and out-of-sample asymptotic optimality for the proposed methods, together with consistency of the estimated adaptive weight function under regularity conditions. The framework applies directly to Euclidean responses and extends to unstructured data by combining our formulation with fixed representation maps. Across a broad set of simulations and real-data studies spanning tabular, image, and text modalities, MoEMA generally improves over competing baselines, demonstrating the effectiveness of our proposed methods.

stat.ML↗

arXiv2026

MUSON: A Reasoning-oriented Multimodal Dataset for Socially Compliant Navigation in Urban Environments

Socially compliant navigation requires structured reasoning about dynamic pedestrians and physical constraints to ensure safe and interpretable decisions. Vision-language models (VLMs) provide a promising foundation for this task because they can integrate visual observations with language-based social knowledge. However, existing untuned VLMs still struggle to reliably understand fine-grained social norms, making task-specific fine-tuning essential. At the same time, no large-scale egocentric dataset is available this task. To address these challenges, we introduce MUSON, a multimodal dataset for short-horizon social navigation containing 10,110 egocentric samples collected across diverse indoor and outdoor social scenes. MUSON adopts a structured five-step chain-of-thought annotation framework comprising perception, prediction, reasoning, action, and explanation. It explicitly models static physical constraints and employs a standardized six-action decision space. Compared with existing social-navigation datasets, MUSON provides consistent annotations for reasoning, actions, and explanations. We evaluate ten representative small-to-medium VLMs on MUSON. Qwen3-VL-8B achieves the strongest decision-level performance, attaining the highest action accuracy of 0.7765 and Macro-F1 score of 0.7490, as well as the lowest collision rate of 0.0609. These results demonstrate that MUSON is an effective and reusable benchmark for advancing socially compliant navigation. The dataset is publicly available at https://github.com/MUSON-dataset/MUSON/releases/tag/v1.0.

cs.CV↗

arXiv2026

Double Descent, Ensemble Emergence, and Large Model Averaging in High-Dimensional Multimodel Prediction

This paper investigates the predictive performance of high-dimensional multimodel prediction, where the number of regressors is comparable to the sample size. Leveraging tools from random matrix theory, we derive the exact limiting out-of-sample risk under a nested model setting and comprehensively characterize the risk landscape. This limiting risk helps to reveal two phenomena: simple weighting inherits the double descent trajectory and its associated variance explosion near the interpolation boundary; strategic weighting triggers an ensemble emergence that suppresses the localized risk surge and yields a globally flat risk surface. Building on this limiting risk, we also propose the Large Model Averaging (LaMA) method, in which we consider the discrepancy between in-sample and out-of-sample risks in the high-dimensional regime. Numerical studies and real data applications confirm that LaMA achieves superior predictive accuracy in high-dimensional environments.

stat.ME↗

arXiv2026

Knowledge-Based Pull Requests: A Trusted Workflow for Agent-Mediated Knowledge Collaboration

AI coding agents are changing the bottleneck in software collaboration: code is increasingly cheap, while understanding intent, negotiating scope, and governing long-term project responsibility remain costly. This paper proposes \emph{Knowledge-Based Pull Requests} (KPR), a trusted workflow for agent-mediated software collaboration across trust boundaries, including open source, enterprise, vendor, contractor, and customer-driven settings. In KPR, an external collaborator's local code, tests, and cleaned agent interaction trace are treated as knowledge sources rather than as the default merge candidate. Agents distill these sources into a human-confirmed knowledge package and render it into reviewer-facing forms such as design memos, risk checklists, test plans, or implementation briefs. A project-owned inner trusted coding agent then regenerates candidate code inside the receiving project's environment under repository context, engineering conventions, tests, and security policy. KPR therefore separates two decisions that traditional pull requests often collapse: whether the knowledge should enter the project, and whether a particular implementation should be merged. We contribute the KPR workflow, a candidate artifact schema, a cost-accounting view, a collaboration gateway architecture, a minimal controlled simulation pilot over seven merged public pull requests, and an evaluation agenda. The pilot shows that KPR packages can be instantiated from real PR material and stress-tested under description ablation, diff ablation, and synthetic poisoned-patch conditions. We position KPR as an empirically testable workflow: its value depends on whether auditable extraction, transformation, and project-side regeneration reduce the cost of understanding and reworking high-context external changes.

cs.SE↗

arXiv2026

Multi-Source Prediction-Powered Inference

Prediction-powered inference integrates a small gold-standard dataset with large pseudo-labeled data, whose labels are generated by machine learning methods, to enhance statistical inference. In modern applications, multiple data sources and diverse machine learning methods often give rise to multiple pseudo-labeled datasets, each encoding potentially different aspects of the underlying information. However, how to optimally combine multiple data sources and machine learning methods for statistical inference remains unclear. To address this problem, we propose a multi-source prediction-powered inference method by aggregating multiple pseudo-labeled datasets together, where the aggregation weights are estimated by minimizing the asymptotic volume of the resulting confidence region. We study both homogeneous settings, where the source and target distributions coincide, and heterogeneous settings, where distributional discrepancies arise between source and target distributions, including covariate shift and domain shift. Theoretically, we establish the asymptotic normality of the proposed estimator and show that the resulting confidence-region volume is asymptotically equivalent to the oracle optimal volume within the proposed weighting class. We further characterize when our method yields smaller confidence regions compared with both classical target-only inference and single-source prediction-powered inference. Simulation studies and a real-data application on dual-energy X-ray absorptiometry measured high body fat prevalence show that MPPI can reduce confidence-region volume while maintaining inferential validity in the settings considered.

stat.ME↗

arXiv2026

Local-GS: Accelerating 3D Gaussian Splatting via Tile-Local Warp Coherence

3D Gaussian Splatting (3DGS) has significantly advanced real-time novel view synthesis by representing scenes as dense collections of anisotropic 3D Gaussian primitives. However, the irregular spatial distribution of Gaussians often leads to poor GPU utilization, as warp divergence and redundant computation degrade rendering performance. To address this, we present Local-GS, a warp-coherent rendering paradigm that, organizes Gaussian primitives with respect to SIMT (Single Instruction, Multiple Threads) execution boundaries rather than scene geometry. Specifically, we propose three warp-coherent stages: a hoisting stage that precomputes shared parameters at tile level, a culling stage that discards warps with no contribution, and a blending stage that replaces per-pixel branching with a uniform instruction stream. Across extensive benchmarks on multiple datasets, Local-GS improves efficiency without compromising quality. As a plug-and-play optimization, it provides additional performance gains to all tested baselines, culminating in a $7.76\times$ speedup on Deep Blending scenes.

cs.CV↗

arXiv2026

SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

Agent Skills are structured packages of procedural knowledge that augment large language model (LLM) agents at inference time. Despite rapid adoption, there is no standard way to measure whether they actually help. We present SkillsBench, a benchmark whose current inventory contains 87 tasks across 8 domains paired with curated Skills and deterministic verifiers. Our latest aggregate evaluation runs the 87-task benchmark under matched no-Skills and curated-Skills conditions for 18 model-harness configurations. Curated Skills raise the average pass rate from 33.9% to 50.5% (+16.6 percentage points; 25.5% normalized gain), with configuration-level gains ranging from +4.1 to +25.7 pp. Focused Skills with at most three modules outperform larger or exhaustive bundles, and smaller models with Skills can match larger models without them. SkillsBench establishes paired evaluation as the foundation for rigorous measurement of Skill efficacy on agentic, expertise-heavy work.

cs.AI↗

arXiv2026

MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling

We present MaxProof, a population-level test-time scaling framework for competition-level mathematical proof in the MiniMax-M3 series. M3 first trains three proof-oriented capabilities -- proof generation, proof verification, and critique-conditioned proof repair -- using a defense-in-depth generative verifier engineered for low false-positive rate. These capabilities are merged into a single released M3 model. At test time, MaxProof treats the model as a generator, verifier, refiner, and ranker, searches over a population of candidate proofs, and returns one final proof through tournament selection. With MaxProof test-time scaling, the M3 model reaches 35/42 on IMO 2025 and 36/42 on USAMO 2026, exceeding the human gold-medal threshold on both.

cs.LG↗

arXiv2026

Glove2Hand: Synthesizing Natural Hand-Object Interaction from Multi-Modal Sensing Gloves

Understanding hand-object interaction (HOI) is fundamental to computer vision, robotics, and AR/VR. However, conventional hand videos often lack essential physical information such as contact forces and motion signals, and are prone to frequent occlusions. To address the challenges, we present Glove2Hand, a framework that translates multi-modal sensing glove HOI videos into photorealistic bare hands, while faithfully preserving the underlying physical interaction dynamics. We introduce a novel 3D Gaussian hand model that ensures temporal rendering consistency. The rendered hand is seamlessly integrated into the scene using a diffusion-based hand restorer, which effectively handles complex hand-object interactions and non-rigid deformations. Leveraging Glove2Hand, we create HandSense, the first multi-modal HOI dataset featuring glove-to-hand videos with synchronized tactile and IMU signals. We demonstrate that HandSense significantly enhances downstream bare-hand applications, including video-based contact estimation and hand tracking under severe occlusion.

cs.CV↗

arXiv2026

Experimental Validation of Skull Acoustic Modelling Strategies for Transcranial Focused Ultrasound Simulation: A Cross-Comparison Study

Accurate acoustic modelling of the skull is essential for simulation-guided transcranial focused ultrasound (tFUS), but commonly used skull parameterisation strategies differ in complexity and reported accuracy. This study experimentally compared five k-Wave skull models: two voxel-wise linear mapping models, two three-layer models, and one single-layer fixed-parameter model. Nineteen regions of interest from five historical and two Thiel-embalmed human skulls were tested at 220 kHz, 680 kHz, and 1000 kHz. Bowl-surface source fields were reconstructed using acoustic holography, and simulated intracranial pressure fields were benchmarked against needle-hydrophone measurements. Across frequencies, mean peak-pressure errors ranged from 20% to 31%, whereas intensity errors reached 41% to 77%. Errors in -6 dB focal volume ranged from 11% to 67%, and focal-position discrepancies were typically several millimetres. Simulations generally predicted smaller insertion losses than measured, indicating a tendency to underestimate skull-related attenuation and overestimate transmitted intracranial exposure. The linear mapping model with fixed attenuation gave the lowest frequency-averaged pressure error, but no model showed a consistent advantage across all metrics. These results show that current skull models can reproduce gross intracranial beam patterns while retaining substantial quantitative uncertainty in exposure, focal coverage, and target localisation.

physics.med-ph↗

arXiv2026

Dual-Cluster Memory Agent: Resolving Multi-Paradigm Ambiguity in Optimization Problem Solving

Large Language Models (LLMs) often struggle with structural ambiguity in optimization problems, where a single problem admits multiple related but conflicting modeling paradigms, hindering effective solution generation. To address this, we propose Dual-Cluster Memory Agent (DCM-Agent) to enhance performance by leveraging historical solutions in a training-free manner. Central to this is Dual-Cluster Memory Construction. This agent assigns historical solutions to modeling and coding clusters, then distills each cluster's content into three structured types: Approach, Checklist, and Pitfall. This process derives generalizable guidance knowledge. Furthermore, this agent introduces Memory-augmented Inference to dynamically navigate solution paths, detect and repair errors, and adaptively switch reasoning paths with structured knowledge. The experiments across seven optimization benchmarks demonstrate that DCM-Agent achieves an average performance improvement of 11%- 21%. Notably, our analysis reveals a ``knowledge inheritance'' phenomenon: memory constructed by larger models can guide smaller models toward superior performance, highlighting the framework's scalability and efficiency.

cs.CL↗

arXiv2026

Semi-Supervised Hyperbolic Hierarchical Clustering with Set-Level Structural Priors

Semi-supervised hierarchical clustering aims to learn a tree structure consistent with data patterns and user-provided supervision. Supervision is usually given as leaf-level relations, such as pairwise must-link/cannot-link constraints or triplet-wise must-link-before constraints. Although useful for regulating local sample relations, such supervision does not directly indicate which samples should form coherent subtrees. Consequently, the non-leaf structure of the learned tree may deviate from the hierarchical organization preferred by ground-truth labels. To address this limitation, we propose a semi-supervised hyperbolic hierarchical clustering method with set-level structural priors. The main contribution is to introduce sets as basic modeling units for hierarchy learning. Each set denotes samples expected to cohere within a subtree and is induced from leaf-level supervision together with a learned constraint-consistent similarity structure. These sets act as soft structural priors for subtree-level supervision, allowing supervision to guide non-leaf hierarchy formation beyond local leaf-level relations. Specifically, we first learn constraint-consistent embeddings to obtain a reliable set partition, then construct constraint-induced sets and estimate inter-set similarities to form set-level structural priors. Finally, these priors are incorporated into a hyperbolic hierarchy objective for continuous tree optimization. Experiments on eleven benchmark datasets and ablation studies show that the proposed method consistently improves label consistency over representative hierarchical clustering baselines while also enhancing similarity-based tree quality.

cs.LG↗

arXiv2026

Reliable Self-Improvement Training by Verifying Reasoning, Not Just Answers

Self-improvement training, where models learn from self-generated solutions, promises sustained capability gains but suffers from a pervasive failure mode: across multiple rounds, compounding reasoning errors cause accuracy to stall or degrade. We trace this drift to standard filtering criteria that retain solutions based solely on final answer correctness, which lets lucky guesses (correct answers with flawed reasoning) contaminate the training data. We propose Verified Self-Improvement (VSI), a framework that conditions data retention on step-level structural integrity rather than just the final output. VSI validates solutions by recomputing arithmetic steps via a computer-algebra library (sympy), checking intermediate consistency, and enforcing domain constraints. Evaluating VSI on GSM8K with Qwen3-4B-Thinking across 5 rounds of self-improvement against four baselines (no verification, outcome verification, majority voting, and VSI with DPO) shows that VSI rejects approximately 34% of correct-answer solutions, successfully isolating lucky guesses. This cleaner training signal drives sustained accuracy gains across all rounds (80.5% to 91.0%), whereas outcome verification plateaus and unverified training collapses. Finally, converting VSI checks into DPO preference pairs trains the model to distinguish sound reasoning from lucky answers, boosting reward accuracy from 46% to 63%. VSI offers a simple, reproducible recipe for robust self-improvement whenever automated reasoning checks are available.

cs.AI↗

arXiv2026

Identification and Inference for Structural Accelerated Failure Time Models via Instrument Interactions

We study causal inference for time-to-event outcomes under right censoring in the presence of unmeasured confounding. Focusing on structural accelerated failure time models, we develop an identification and inference framework that exploits interactions among instrumental variables. The proposed approach does not rely on classical instrumental variable validity and yields valid causal inference under both valid and invalid instruments, provided that the interaction-based identification condition holds. To accommodate right censoring, we construct a censoring-adjusted observed data moment function using an augmented inverse probability censoring weighting approach. The resulting moment function is Neyman orthogonal with respect to nuisance functions and enjoys a double robustness property, enabling valid inference under flexible nuisance estimation. Estimation and inference are conducted using generalized empirical likelihood, which is well suited to settings with many potentially weak interaction-based moment conditions. We establish consistency, and asymptotic normality under many weak moment asymptotics, and develop diagnostic tools to assess interaction-based identification strength and overidentifying restrictions. Simulation studies demonstrate favorable finite sample performance across a range of censoring rates and instrument configurations. An application to UK Biobank data illustrates the practical relevance of the proposed method for causal survival analysis in large-scale observational studies.

stat.ME↗

arXiv2026

SchGen: PCB Schematic Generation with Semantic-Grounded Code Representations

Printed circuit board (PCB) schematic design defines nearly all electronic hardware, but it remains manual and expertise-intensive. While generative AI has advanced digital and analog IC design, PCB schematic generation from natural-language intent is largely unexplored. This paper presents SchGen, the first large language model that generates editable PCB schematics from natural-language requests. The key challenge lies in the lack of an LLM-suited representation and a large-scale dataset. Current schematic formats are dominated by verbose, tool-specific syntax and geometry-heavy descriptions, making them difficult to generate reliably. We introduce a semantically grounded code representation that encodes schematic editing primitives with relative placement and pin-name-based wiring, transforming a geometry-driven generation problem into a semantics-driven matching task amenable to LLMs. We further construct a large-scale dataset of PCB schematics paired with user prompts via a human-agent collaborative pipeline that converts open-source hardware designs into our representation. Experiments show that SchGen significantly outperforms alternative representations and even larger general-purpose LLMs on wire connectivity accuracy and functional correctness. Our results highlight the critical role of representation design in enabling generative models for complex hardware design tasks.

cs.AI↗