Searcharxiv⌕ Search

arXiv subjects

Wenbo Li

Publications and source records attributed to Wenbo Li.

At least 37 records · Page 2Linked to original sources
arXiv2026

Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks

Merging low-rank adapters (LoRAs) promises to eliminate the overhead of swapping task-specific weights at inference time. However, existing merging methods assume every layer needs the same rank budget. Further, some methods assume that rank budget needs to be split equally among the tasks too. We show this uniform-budget assumption is a major source of the performance gap between merged and per-task LoRAs. However, rank selection is an NP hard problem. To this end, we introduce Net Utility, a data free metric that first decomposes every task LoRA by its Singular Value Decomposition (SVD) and scores each of those singular directions by its task utility and its interference with other tasks directions. Next, we globally pool these scores to select singular directions with the highest values with a constraint on the total number of directions selected. The proposed Net Utility metric is applied on top of five different merging methods across three different merging spaces. The merging is done over two sets of tasks, vision and language tasks. Net utility based rank allocation outperforms its counterparts without that allocation. On average, over vision tasks it achieves +2.1% improvement in performance, and +2.2% improvement over the language tasks.

cs.LG↗
arXiv2026

Condensed PIPG Sequential Convex Optimization for Reusable-Rocket Powered Landing with Strong Aerodynamics

Reusable-rocket powered landing under strong aerodynamics couples variable mass, free final time, and bounded aerodynamic controls through nonlinear velocity-frame dynamics. This paper develops a condensed proportional--integral projected-gradient (PIPG) sequential-convex method whose principal contribution is an exact reduced-space inner architecture. Because the problem contains only six terminal hard equalities and no state path constraints, 217 nodal-state variables and 210 trapezoidal dynamics equalities are eliminated from the 31-node convex subproblem, leaving 101 primal variables and six terminal equalities. Row-orthogonal preconditioning, fixed-size matrix--vector products, and nodewise circular-epigraph projections then yield a customized PIPG kernel. Physical consistency of the angle-dependent axial force is maintained by gradually releasing drag sensitivity between the reference squared angle and an epigraph variable $A$, together with a convex tightness term. A pointwise Hamiltonian argument shows that the fully released limiting subproblem admits a tight optimum satisfying $A=α^2+β^2$. Deterministic annealing, two-stage inner accuracy, and a rejected-on-failure threefold extrapolation are secondary outer-loop accelerators.

math.OC↗
arXiv2026

ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training

Robotic manipulation faces a fundamental scaling challenge: robust generalization demands broad physical experience, yet action-labeled robot trajectories are expensive to collect and inherently limited in diversity. Egocentric videos offer a far more scalable source of embodied experience, capturing object interactions, contact dynamics, tool use, and long-horizon behaviors across diverse environments. The central challenge is how to convert this abundant but action-free experience into effective robot control. We introduce ZimaBlue, a scalable framework for learning generalizable World Action Models (WAMs) from large-scale video. ZimaBlue follows a three-stage training curriculum: it first performs causal embodied video pre-training on large-scale human and robot egocentric videos, then grounds the learned visual dynamics in heterogeneous robot trajectories through video-action mid-training with a unified action representation, and finally specializes the model to a target robot for deployment. To make generative WAMs practical for real-time control, ZimaBluefurther adopts an asynchronous Slow-Fast dual-system architecture, where a high-capacity Slow world model provides generalizable spatiotemporal representations and a lightweight Fast branch enables 30 Hz action prediction on NVIDIA RTX 4090. On real-robot zero-shot evaluations, scaling from target-robot data alone to over 120,000 hours of embodied video improves success from 36.1% to 77.8%. ZimaBlue further delivers strong performance across multiple benchmarks, with particularly pronounced gains on unseen tasks.

cs.CV↗
arXiv2026

RubricRM: Generative Reward Modeling via Dynamic Rubrics for Image Generation and Editing

Reward models play an essential role in aligning visual generative models, yet most existing visual reward models use a single scalar score or rely on fixed criteria that cannot adapt to different instructions. This limits both interpretability and task sensitivity, especially for text-to-image generation and instruction-based image editing, where different inputs require different evaluation dimensions. We propose RubricRM, a pairwise generative reward modeling framework that first produces an input-specific rubric with evaluation dimensions, weights, and scoring criteria, and then applies the rubric to score candidate images. We train dedicated RubricRM models for text-to-image generation and image editing using a two-stage training pipeline: supervised fine-tuning teaches the model the rubric-based scoring paradigm, while GRPO further improves scoring through fine-grained dimension-level rewards. Experiments on multiple generation and editing benchmarks show that RubricRM outperforms existing specialized reward models and remains competitive with strong proprietary MLLM judges despite using smaller backbones. Our models, data, and code are available at https://github.com/zijiankan/RubricRM.

cs.CV↗
arXiv2026

SlidesGen-Bench: Evaluating Slides Generation via Computational and Quantitative Metrics

The rapid evolution of Large Language Models (LLMs) has fostered diverse paradigms for automated slide generation, ranging from code-driven layouts to image-centric synthesis. However, evaluating these heterogeneous systems remains challenging, as existing protocols often struggle to provide comparable scores across architectures or rely on uncalibrated judgments. In this paper, we introduce SlidesGen-Bench, a benchmark designed to evaluate slide generation through a lens of three core principles: universality, quantification, and reliability. First, to establish a unified evaluation framework, we ground our analysis in the visual domain, treating terminal outputs as renderings to remain agnostic to the underlying generation method. Second, we propose a computational approach that quantitatively assesses slides across three distinct dimensions - Content, Aesthetics, and Editability - offering reproducible metrics where prior works relied on subjective or reference-dependent proxies. Finally, to ensure high correlation with human preference, we construct the Slides-Align1.5k dataset, a human preference aligned dataset covering slides from nine mainstream generation systems across seven scenarios. Our experiments demonstrate that SlidesGen-Bench achieves a higher degree of alignment with human judgment than existing evaluation pipelines. Our code and data are available at https://github.com/YunqiaoYang/SlidesGen-Bench.

cs.CL↗
arXiv2026

Dual-Thrust Switching Analytical Guidance Algorithm for Powered Landing with Attitude Smoothness Optimization

Traditional numerical guidance methods for powered landing of reusable rockets are typically constrained by high computational complexity and inadequate real-time performance. Moreover, insufficient consideration of attitude smoothness often induces severe fluctuations in control commands; meanwhile, most existing approaches are tailored for single-thrust scenarios, failing to accommodate the guidance requirements of multi-engine thrust switching. To mitigate these limitations, this paper proposes an analytical guidance method optimized for attitude smoothness, which supports dual-thrust-mode switching. First, a corresponding optimal control problem is formulated, and it is theoretically proven that the optimal attitude command takes a concise piecewise cubic function form. This transforms complex trajectory optimization into a parametric analytical optimization problem, yielding a substantial improvement in computational efficiency. Further, a three-phase guidance framework is designed to enable adaptive determination of the guidance activation point and thrust switching point; when integrated with an aerodynamic correction strategy, this framework enhances the method's adaptability in complex flight environments particularly under high lift-to-drag ratio conditions. Simulation results demonstrate that the attitude command profile generated by the proposed method aligns closely with the theoretical optimal solution, with an ultra-short computation time, confirming its strong potential for online real-time implementation. Even under stringent conditions (e.g., limited thrust adjustment range, high lift-to-drag ratios, and parameter deviations), the method consistently achieves high-precision landing, showcasing promising prospects for engineering applications.

math.OC↗
arXiv2026

PhysMaster: Building an Autonomous AI Physicist for Theoretical and Computational Physics Research

Advances in LLM reasoning and tool use have enabled agentic science, yet frontier theoretical and computational physics remains challenging because research requires deep domain expertise, long-horizon reasoning, and reliable numerical computation. We introduce PRL-Bench, a research-reproduction benchmark adapted from 100 Physical Review Letters papers across major areas of modern physics. PRL-Bench distills realistic research workflows into traceable tasks with explicit intermediate artifacts and diverse evaluation rubrics; each task is estimated by domain experts to require more than six hours for a specialized PhD student to reproduce independently. Evaluations show that existing agents remain unreliable on extended research workflows. We therefore present PhysMaster, a scientific agent combining adaptive MCTS-based multi-trajectory exploration with hierarchical memory to improve long-horizon robustness and knowledge accumulation. PhysMaster achieves the highest overall PRL-Bench score of 51.08, outperforming Codex, OpenHands, OpenClaw, and ReAct, and yields relative improvements of 14.13 percent to 93.38 percent across backbone models. Error analysis shows that PhysMaster substantially reduces failures from incomplete long-horizon execution, while remaining bottlenecks lie in physics knowledge and analytical reasoning. Together, PRL-Bench and PhysMaster provide a rigorous benchmark and effective system for advancing autonomous AI research in frontier physics.

cs.AI↗
arXiv2026

Submillisecond Sequential Convex Optimization for Powered Landing via Dynamics Condensation and xPIPG

Powered landing with variable mass, free final time, and quadratic aerodynamic drag requires the repeated solution of local convex subproblems, whose main online cost lies in the long dynamics-equality chain and the inner iterations. This paper develops a condensed sequential convex approximation designed for low latency. Exact block elimination removes 217 intermediate-state components and 210 interval equations from a 31-node model, leaving 100 primal variables coupled by six terminal equalities. A low-weight energy term and fixed quadratic proximal regularization make the ideal surrogate strongly convex with predictable curvature. The inner solver is an extrapolated proportional--integral projected gradient (xPIPG) implemented with fixed-size arrays, $3\times3$ interval solves, and a fused one-pass node map. The one-pass map is a deliberate low-cost approximation, not the exact joint proximal operator. We therefore evaluate the timed code by nonlinear trajectory residuals and independent physical checks rather than by a claim of exact KKT convergence. The single-precision C implementation completes one plan in four outer updates and 336 xPIPG updates. On an Intel Core i7-10875H, the median end-to-end solve time is \SI{374}{\micro\second} and the P99 value is \SI{512}{\micro\second}. All 100 common initial-state perturbations pass validation, and the median remains below \SI{0.7}{\milli\second} for 15--51 nodes. An independent high-accuracy first-order-hold integration gives a terminal position error of \SI{0.183}{\meter}. Within the stated model, hardware, stopping rule, and timing boundary, this is, to the authors' knowledge, the first submillisecond end-to-end sequential-convex solve for a single powered-landing trajectory.

math.OC↗
arXiv2026

Teacher-Feature Drifting: One-Step Diffusion Distillation with Pretrained Diffusion Representations

Sampling from pretrained diffusion and flow-matching models typically requires many forward passes to generate diverse and high-fidelity images. Existing distillation methods often rely on multiple auxiliary networks, carefully designed training stages, or complex optimization pipelines. In this work, we revisit the recently proposed Drifting Model objective and show that a single drifting loss can be directly used to simplify one step distillation. A key observation is that the pretrained diffusion teacher itself already provides a strong representation space. Unlike the original Drifting Model, which relies on an additional pretrained feature extractor, we use intermediate hidden states of the pretrained teacher model as the feature representation. This removes the need for training or introducing an extra representation network while preserving a semantically meaningful feature geometry for drifting. Furthermore, we introduce a lightweight mode coverage loss to mitigate mode collapse during distillation and encourage the student generator to cover diverse teacher-supported regions. Extensive experiments on ImageNet and SDXL demonstrate that our method achieves efficient one step generation with competitive image quality and diversity, achieving FID scores of 1.58 on ImageNet-64$\times$64 and 18.4 on SDXL, while substantially simplifying the overall distillation framework.

cs.CV↗
arXiv2026

HalluScope: Fine-grained Hallucination Diagnosis for Multimodal Large Language Models

Although Multimodal Large Language Models have achieved strong performance across a wide range of vision-language tasks, they still suffer from hallucinations, where model outputs become inconsistent with the visual content, textual context, or commonsense knowledge. Existing studies primarily address this problem through coarse-grained detection. However, these approaches often provide insufficient diagnostic information for understanding hallucination types and supporting downstream hallucination mitigation. To bridge this gap, we propose fine-grained hallucination diagnosis for MLLMs, a new unified task that jointly performs hallucination detection, classification, and interpretable explanation generation. We develop an automated data generation pipeline and construct HalluScope-30K, a large-scale diagnostic dataset covering eight sources and five task categories. Based on this dataset, we design a multi-granular joint reward function and train two diagnosis models, HalluScope-4B and HalluScope-8B, which achieve state-of-the-art performance on both the MHALO benchmark and our fine-grained hallucination classification benchmark. Notably, detection and classification are mutually beneficial under joint optimization. Furthermore, diagnosis-driven feedback experiments show that the fine-grained diagnostic explanations produced by our model effectively guide target models to correct their hallucinations, with full diagnosis substantially outperforming all baselines on both Qwen3-VL-8B-Instruct and LLaVA-1.5-7B. Our code, data, and models are available at https://github.com/wkinglin/HalluScope.

cs.CV↗
arXiv2026

JoyAI-Video-Edit: Real-Time Open-Ended Video Editing with Autoregressive Diffusion

Real-time video editing requires low-latency causal generation with bounded computational resources while preserving source fidelity and long-term temporal consistency. We present JoyAI-Video-Edit, a 16B-parameter autoregressive diffusion framework for real-time, open-ended video editing without access to future frames or a predefined video duration. Our method combines chunk-wise autoregressive adaptation, Source-Anchored Distribution Matching Distillation (SA-DMD), and Long-Horizon Autoregressive Distillation to reduce train--inference mismatch, preserve source fidelity during two-step generation, and mitigate accumulated temporal drift. Extensive automatic and human evaluations show that JoyAI-Video-Edit substantially outperforms existing streaming editors and remains competitive with strong offline systems on both short and long videos. The complete system achieves end-to-end 720p video editing at approximately 30 FPS on a single Nvidia B200 GPU. Code is available at https://github.com/jd-opensource/JoyAI-Video-Edit.

cs.CV↗
arXiv2026

PROGRESS: Coverage-guided RL to Train Search-augmented LLM Agent

Existing search-augmented LLM agents are trained using Reinforcement Learning to boost its reasoning capabilities. However, these approaches primarily rely on outcome-level rewards, which provide little supervision over search behavior and overlook agent's ability to decompose complex queries properly. To mitigate this issue, we propose PROGRESS which utilizes teacher-guided coverage reward to explicitly shape decomposed query generation of the policy model. During training, frozen teacher models are used to decompose complex queries into essential search queries. These essential search queries are utilized to guide the search behavior of the policy model. Integrated into an R1-style training framework, our approach provides lightweight guidance over query decomposition decisions without dense process-level supervision. Experiments show that coverage-guided RL improves overall task performance, highlighting the importance of explicitly supervising search strategies in agentic LLMs.

cs.AI↗
arXiv2026

Role-Break in Attention Heads: Understanding and Detecting Hallucinations in VLMs

Despite remarkable progress in vision-language generation, Vision-Language Models (VLMs) remain prone to hallucinations, producing content that is inconsistent with or unsupported by the input image. Existing works largely design detection or mitigation methods around one specific hallucination pattern, such as visual-textual imbalance, but real VLM hallucinations arise from a mixture of multiple patterns, so signals bound to a single pattern struggle to remain stable across models and tasks. Under a unified head-level view, we find that hallucination-induced changes manifest as localized deviations from each head's faithful contextual behavior, a phenomenon we term Role-Break. Detailed analysis reveals that these deviations are systematically organized across attention heads, contextual sources, and deviation directions, and that the resulting signal is linearly readable once head identity is preserved. Based on these findings, we build a lightweight linear detector on top of Role-Break that requires no fine-tuning of the VLM, whose feature dimension stays below 5,000 and reaches an average AUROC of 93.23 across six VLMs and four benchmarks. A small-scale intervention experiment further shows that the detected tokens can be directly acted upon in the discriminative setting.

cs.CV↗
arXiv2026

BioPro: Towards Difference-Aware Gender Fairness for Vision-Language Models

Vision-Language Models (VLMs) inherit significant social biases from their training data, notably in gender representation. Current fairness interventions often adopt a difference-unaware perspective that enforces uniform treatment across demographic groups. These approaches, however, fail to distinguish between contexts where neutrality is required and those where group-specific attributes are legitimate and must be preserved. Building upon recent advances in difference-aware fairness for text-only models, we extend this concept to the multimodal domain and formalize the problem of difference-aware gender fairness for image captioning and text-to-image generation. We advocate for selective debiasing, which aims to mitigate unwanted bias in neutral contexts while preserving valid distinctions in explicit ones. To achieve this, we propose BioPro (Bias Orthogonal Projection), an entirely training-free framework. BioPro identifies a low-dimensional gender-variation subspace through counterfactual embeddings and applies projection to selectively neutralize gender-related information. Experiments show that BioPro effectively reduces gender bias in neutral cases while maintaining gender faithfulness in explicit ones, thus providing a promising direction toward achieving selective fairness in VLMs. Beyond gender bias, we further demonstrate that BioPro can effectively generalize to continuous bias variables, such as scene brightness, highlighting its broader applicability.

cs.AI↗
arXiv2026

Microsecond-Class Powered-Descent Optimization via Exact Condensation and Strong Convex Regularization

Fuel-dominant powered descent can be written as a convex program, but the usual full-state epigraph formulation still carries many state variables, fuel epigraph variables, and dynamics equalities. In addition, the pure-fuel objective provides no strong-convexity curvature. This paper combines three structural reductions. First, a dimensionally consistent low-weight energy term makes the control solution unique. Second, a terminal-state sensitivity recursion eliminates every intermediate state exactly; invertible row normalization turns the 30-node baseline with 300 primal variables and 174 equality multipliers into a problem with 90 control variables and six terminal multipliers. Third, the shared radial structure of the fuel norm and thrust ball gives an exact closed-form proximal operator consisting of group shrinkage followed by magnitude clipping. The condensed problem is solved with a fixed-budget extrapolated proportional--integral projected-gradient iteration implemented in fixed-size C17 arrays. In a Mars powered-descent case with energy weight 0.02 and relative reference-solution tolerance $10^{-3}$, the iteration count decreases from 2744 for the pure-fuel full-state epigraph baseline to 93, while the fuel metric increases by only 0.033\%. The mean end-to-end solve time is \SI{68.2}{\micro\second}, and P99 is \SI{128.1}{\micro\second}, on an Intel i7-10875H. Because the thrust set in this test case is already a convex ball, the contribution is fast solution of the convex core rather than a new lossless-convexification theorem.

math.OC↗
arXiv2026

Geo3R: Mitigating Spatial Reasoning Hallucination in Multimodal Large Language Models

Despite remarkable progress in visual understanding, Multimodal Large Language Models (MLLMs) remain prone to hallucinations when reasoning about spatial relationships, often producing judgments that contradict the true 3D structure of the scene. Though several existing works have proposed to mitigate hallucinations, our analysis indicates that they show limited effectiveness in spatial reasoning, as they fail to bridge the fundamental gap between 2D visual representations and 3D spatial reality. Based on this finding, we define hallucinations arising from insufficient spatial structure modeling as spatial reasoning hallucination, a subcategory of relation hallucination that existing mitigation methods fail to address. We further identify three typical scenarios where such hallucinations frequently occur: perspective effects, object orientation, and viewpoint changes. To this end, we propose Geo3R, a training-free, plug-and-play framework that incorporates geometric evidence and structured 3D reasoning to mitigate spatial reasoning hallucination. Experiments on three benchmarks, covering 18 tasks across all three scenarios, show that Geo3R substantially reduces spatial reasoning hallucination across diverse MLLMs without additional training, outperforming existing models and methods.

cs.CV↗
arXiv2026

Do LLMs Ask the Right Questions? Evaluating GPT-Generated Surveys as Instruments for Measuring Social Attitudes

Understanding human beliefs and social attitudes often relies on carefully designed survey instruments. Recent work has suggested that large language models (LLMs) could automate parts of this process by generating surveys at scale, raising questions about the comparability of such instruments to literature-grounded, human-designed surveys. We present a controlled empirical comparison between GPT-generated surveys and established survey baselines across three social domains: climate change, immigration, and diversity, equity, and inclusion (DEI). GPT-generated surveys were produced using a fixed prompting framework enforcing a 3x3 structure over beliefs, perceptions, and behaviors, while human baselines were assembled from validated instruments to match survey length and construct coverage. We collected responses from U.S.-based participants, who completed both survey types, allowing direct within-subject comparison. We analyze differences in response distributions, clustering behavior, and alignment with self-identified stances. Our results show that GPT-generated surveys capture the same dominant attitudinal divisions as human-designed instruments, while exhibiting differences in the resolution of belief structure and group separation. These findings suggest that LLM-generated surveys are suited for exploratory and large-scale analyses, and can be used to complement expert-designed instruments.

cs.CY↗
arXiv2026

A Dual-Bearing Magnetorheological Grease Clutch with Intention-Based Demagnetization for Wearable Haptic Feedback

This paper presents the design, modeling, and control of a dual-bearing magnetorheological grease (MRG) clutch for wearable haptic feedback. Compared with conventional MR fluid devices, the proposed clutch avoids leakage-related reliability degradation while achieving high torque density in a compact structure. To provide physical insight into the torque-generation mechanism, a physics-inspired interpretive model is introduced to capture the dominant relationship among excitation current, magnetic-field evolution in the bearing gaps, and clutch locking torque. To mitigate the undesirable ``sticky'' sensation caused by passive bidirectional braking, an intention-based control strategy with active demagnetization is further developed to enable smoother release during human--robot interaction. Experimental characterization shows that the proposed clutch achieves a maximum locking torque of 43.42\,N$\cdot$m at 1.3\,A and a torque-to-mass ratio of 96.5\,N$\cdot$m/kg. Bench tests, replay validation, teleoperation experiments, and user studies indicate that the proposed approach accelerates clutch release and improves perceived release transparency and contact-to-release smoothness, while maintaining effective multi-level kinesthetic rendering.

eess.SY↗