arXiv subjects

Kai Tang

Publications and source records attributed to Kai Tang.

At least 19 recordsLinked to original sources

arXiv2026

Learn from the Gap: Differential-Aware Advantage Pruning with Adaptive Rollout Sampling for GRPO

Recently, Group Relative Policy Optimization (GRPO) and its variants have been developed for policy optimization and demonstrated notable performance gains. However, these methods usually incur substantial computational overhead due to per-question multi-rollout sampling and repeated per-token probability evaluation across rollouts. Furthermore, low-information or highly homogeneous trajectories can degrade downstream learning signal efficiency, hindering model optimization and limiting final performance. To address these issues, we propose FastRL, a novel plug-and-play reinforcement learning framework that simultaneously improves training efficiency and the effectiveness of policy learning. Specifically, 1) We introduce an advantage-aware pruning strategy to selectively preserve high-advantage trajectories while maximizing inter-trajectory gradient diversity. 2) Then, we design an adaptive rollout sampling mechanism to dynamically adjust the sampling scale across different training stages based on historical pruning distributions, balancing exploration adequacy and computational efficiency. Experiments demonstrate that FastRL can be seamlessly integrated into GRPO, DAPO, and GSPO variants, achieving an average 2.07$\times$ training speedup on Geometry3K and GeoQA8K-R1V, along with an approximately 1.64\% improvement in average accuracy on visual reasoning benchmarks. Source codes will be available at https://github.com/Nicozwy/FastRL.

cs.AI↗

arXiv2026

LAW: Unifying Language, Action, and World Modeling for Agentic Physical Intelligence

World-action models jointly predict actions and future world states, but long-horizon manipulation further requires language reasoning to plan subtask objectives, assess execution progress, and guide action selection. We present LAW, a closed-loop VLM--WAM framework that unifies language reasoning, action generation, and world modeling through semantic subtask progress. Within the LAW agent, a hybrid-expert vision-language model plans the remaining subtasks, estimates the values of and fuses actions proposed by a world-action model, and asynchronously assesses execution status to support continued execution and subtask transitions. Within the world-action model, Mixture-of-Transformer Experts integrates modality-specific and shared computation for joint video--action modeling. Dual-stream hierarchical visual memory maintains temporal context for both models, while subtask-local value learning aligns action evaluation with semantic objectives and preserves supervision from successful prefixes of failed trajectories. LAW achieves 94.21% average success across 50 RoboTwin tasks. Replacing the WAM backbone with OpenWAM-$alpha$ further raises the average success rate to 95.56%, establishing a new open-source state of the art on RoboTwin. Despite having only 2B parameters, our VLM also outperforms variants using Embodied-R1.5 and GPT-5.5 High. On three long-horizon real-robot tasks, LAW achieves 65% average success, compared with 28.33% for $pi_{0.5}$ and 25% for Fast-WAM.

cs.RO↗

arXiv2026

SafeDojo: Safe Reinforcement Learning for VLA via Interactive World Model

Safe control is a prerequisite for real-world embodied intelligence, for which safe reinforcement learning has emerged as a promising paradigm. However, existing safe reinforcement learning methods either require costly real-world exploration or depend on hand-crafted safety functions. Neither scales to vision-language-action models deployed in open-world physical environments. We propose SafeDojo, a model-based safe reinforcement learning framework that improves vision-language-action policies through world-model imagination. SafeDojo performs online reinforcement learning on top of an interactive video world model. A ResNet success classifier scores predicted future frames, while a lightweight safety head predicts per-step contact probabilities from the current latent context and proposed action chunk, jointly assessing task execution and trajectory safety. These decoupled signals are balanced through Lagrangian-constrained GRPO. On SafeLIBERO Level I, SafeDojo improves average safe-success rate by 8.25 percentage points over the strongest baseline; on the zero-shot Level II setting, it is competitive with the strongest model-free baseline. Real-world Franka deployment further shows the best average task and safe-success rates across five tasks. These results support world-model-based safe reinforcement learning as a promising path toward safer embodied intelligence.

cs.RO↗

arXiv2026

SimOPD: Relational Knowledge Transfer for On-Policy Distillation

A key challenge in on-policy distillation (OPD) is extracting effective supervision from teacher knowledge. Supervision from token probabilities may underutilize internal knowledge, while direct hidden-state matching forces the student to shift to the teacher's feature space and complicates cross-architecture transfer. Inspired by contrastive learning's use of relationships between instances, we propose \textbf{SimOPD} to transfer teacher internal knowledge through response similarities. Specifically, SimOPD uses paired memory banks of student-generated responses to extract supervision from the teacher's internal representations as similarity distributions. The student then matches these distributions within its own representation space, retaining representational freedom and removing the need for cross-model projection. Extensive experiments show that SimOPD improves the overall average over the strongest baselines by up to 2.6 and 3.1 percentage points in same- and cross-architecture settings.

cs.AI↗

arXiv2026

CoCA: Criterion-Aware Credit Assignment for Rubric-Based Reinforcement Learning

Rubric-based reinforcement learning (RL) evaluates responses against explicit criteria to construct rewards, offering a viable way to train large language models on open-ended tasks. However, standard methods compress criterion-level feedback into scalar rewards and assign uniform token advantages, leaving structured rubric supervision underutilized. Recent work seeks finer-grained guidance by treating rubrics as privileged information to derive self-distillation targets, but the resulting learning signal is not grounded in criterion outcomes and can lead to reward degradation. This motivates us to reframe the problem as criterion-aware credit assignment, where privileged rubric information serves as a relevance cue for credit allocation rather than a source of imitation targets. Accordingly, we propose CoCA, which derives criterion-level credits from group-relative outcomes and allocates them across tokens based on criterion-induced distribution shifts, yielding both outcome-grounded and fine-grained learning signals. Empirically, CoCA achieves an average gain of +4.2% over standard rubric-based RL across medicine and science benchmarks, demonstrating more effective learning on open-ended tasks.

cs.LG↗

arXiv2026

SpikeOcc: Towards Resilient Multimodal 3D Occupancy Perception via Spiking Experts

Multimodal 3D occupancy prediction underpins safe autonomous driving through comprehensive scene understanding. However, existing methods have largely pursued higher accuracy through sophisticated fusion, leaving resilience to missing sensors underexplored. In this paper, we investigate spiking neural networks (SNNs) as a means of building more resilient multimodal perception. Still, the discrete nature of spiking neurons makes optimization challenging. To overcome this difficulty, SpikeOcc is introduced with two complementary components, a Spiking Mixture-of-Experts (SMoE) and Sensor Availability Learning (SAL). Specifically, SMoE organizes RGB, LiDAR, and joint fusion into specialized pathways, enabling full-width unimodal representations while retaining the original multimodal features via a bounded residual. Beyond sensor dropout, SAL employs a staged curriculum with explicit routing supervision, progressively improving robustness under severe sensing conditions. Remarkably, SAL also yields notable gains for conventional models, highlighting its broader potential for resilient multimodal learning. Extensive experiments demonstrate that SpikeOcc maintains competitive full-sensor accuracy while markedly raising the performance floor under sensor outages. SpikeOcc illustrates the promise of neuromorphic computing, pointing toward more reliable and energy-efficient autonomous driving.

cs.CV↗

arXiv2026

ReMAP: Restoring the Perceptual Cycle with Reasoning-Time Latent Visual Memory

As multimodal large language models (MLLMs) reason for longer, attention to the initial visual input diminishes, weakening visual grounding. Visual memory reintroduces visual evidence during reasoning. We conduct a controlled analysis of visual memory along three axes: curation, organization, and access. We find that local evidence benefits from global context, compact latent representations balance accuracy and visual-context cost, and the utility of memory access depends on the reasoning state. Guided by these findings, we propose ReMAP (Reasoning-Time Memory-Augmented Perception), which couples two complementary latent memories: a static, question-conditioned Global memory that preserves scene and cross-image context, and a dynamic Local memory that uses this context as an anchor while selecting and re-encoding region-level evidence according to the current reasoning state. Both memories return compact latent tokens inserted into the reasoning sequence, and a reinforcement-learning access policy trained with branched rollouts decides when to continue reasoning or invoke Global or Local memory. On ten benchmark families, ReMAP outperforms prior visual-memory methods on all four multi-image benchmarks, exceeding the strongest prior results on MuirBench and MIMIC by 8.38 and 14.84 percentage points. Across four backbone families, enabling memory access improves over the same trained model with memory disabled, and on shared V*Bench, CV-Bench-2D, and MuirBench questions ReMAP reduces the visual tokens entering the reasoning sequence by 51.0-76.8% relative to the native-resolution backbone. Further analyses show that Global and Local memory form distinct yet complementary latent representations. Together, these components restore the perceptual cycle by letting the reasoning state trigger targeted visual retrieval, with the retrieved evidence guiding subsequent reasoning.

cs.CV↗

arXiv2026

Teacher-Student Gaps Are Not Enough: Outcome-Guided On-Policy Distillation for Multi-Turn Autonomous Agents

On-policy distillation (OPD) trains a student on its own trajectories with dense teacher supervision. Recent work on OPD for multi-turn autonomous agents often treats large teacher-student token-level distributional gaps as promising intervention points, linking larger gaps to a greater need for correction. Yet, our empirical analysis reveals a supervision-benefit mismatch: large gaps can be benign, while small gaps can be outcome-critical. Teacher-student gaps capture differences at the current turn, whereas the benefit of teacher guidance depends on how the current student interacts with the environment afterward. The student may still succeed despite choosing an action that differs from the teacher's, while a teacher-preferred action may lead to a state from which the student cannot complete the task. Local gaps alone are therefore not enough to determine whether teacher guidance benefits the current student. Effective supervision should instead emphasize guidance that the current student can translate into better final task outcomes. Accordingly, we propose Outcome-Guided On-Policy Distillation (OG-OPD), which applies trajectory-relative weighting to teacher supervision and calibrates these weights using final task outcomes from paired student continuations. This calibration selectively strengthens supervision on the student's original trajectories at turns where teacher guidance benefits the current student. Across ALFWorld, ScienceWorld, and WebShop, OG-OPD consistently outperforms baselines under diverse settings. It improves task success rates by 3.6-17.7 percentage points over vanilla OPD and by up to 7.0 percentage points over the strongest baseline.

cs.LG↗

arXiv2026

FADE: Mitigating Hallucinations by Reducing Language-Prior Dominance in Large Vision-Language Models

Despite the impressive capabilities of Large Vision-Language Models (LVLMs), they remain susceptible to hallucination, generating content inconsistent with the input image. Recent studies attribute this to the dominance of language priors over visual inputs and employ contrastive decoding methods to mitigate this dominance, but the mechanistic origin remains unexplored. We investigate the information flow through each transformer layer and find that attention modules consistently aggregate visual evidence, while FFN modules at critical layers act as the source of language priors. These priors can override visual evidence, causing correct predictions in intermediate layers to drift toward incorrect outputs. Based on this insight, we propose FADE (FFN Attenuation for DEcoding), a training-free method that attenuates FFN outputs to reduce language-prior dominance. Evaluations on POPE, CHAIR, and MME benchmarks across LLaVA-1.5, mPLUG-Owl2, and InstructBLIP show that FADE effectively mitigates hallucinations while preserving inference efficiency.

cs.AI↗

arXiv2026

Hermitian manifolds with constant holomorphic sectional curvature or real bisectional curvature

An old conjecture in Hermitian geometry states that a compact Hermitian manifold with constant holomorphic sectional curvature is Kähler when the constant is nonzero and Chern flat when the constant is zero. Yang--Zheng introduced the real bisectional curvature as a generalization of the holomorphic sectional curvature and conjectured that, a compact Hermitian manifold with constant real bisectional curvature has zero constant and is Chern flat. They proved that the constant cannot be positive. In this paper, we first prove their conjecture. For a negative constant, we show that the metric is pluriclosed and use comparison with a negative Kähler--Einstein metric to obtain a contradiction. For the zero constant, we combine a torsion identity of Lin--Ren with the Bochner formula of Zhou--Zheng to prove Chern flatness. We also prove that a pluriclosed metric with constant negative holomorphic sectional curvature on a compact Kähler manifold is Kähler.

math.DG↗

arXiv2026

Constant $k$th-mixed curvature on locally conformal Kähler manifolds

In this paper, we consider compact locally conformal Kähler manifolds with constant $k$th-mixed curvature. By using a recent method of Huang-Wan for constant Chern holomorphic sectional curvature, we prove that if a compact LCK manifold has nonzero constant $k$th-mixed curvature, then its Hermitian metric is Kähler. For the second mixed curvature, the same conclusion also holds when the curvature constant is zero. We also study some special parameters related to the general constant mixed curvature conjecture. In particular, we obtain a torsion-energy identity and a sign obstruction for compact Hermitian manifolds, and characterize an exceptional Kähler case by Bochner--Kähler geometry.

math.DG↗

arXiv2026

STAR-Pro: Stage-Wise Token Adaptive Reduction with Progressive Refinement for Efficient Large Vision-Language Models

Large vision-language models (LVLMs) achieve strong multimodal understanding, but the hundreds to thousands of visual tokens they process impose substantial computational overhead, motivating training-free visual token pruning. In this work, we conduct two complementary analyses of visual token pruning. First, we measure the feature-space coverage of tokens retained before cross-modal fusion and find that aggressive pruning discards substantial visual information. Second, we track text-to-visual attention across decoder layers and find that the visual tokens considered important change substantially with depth, making one-shot pruning decisions unreliable. Together, these findings show that effective pruning should preserve broad visual coverage before fusion and progressively refine the retained tokens as cross-modal evidence evolves during fusion. We therefore propose STAR-Pro (STage-Wise Adaptive Token Reduction with Progressive Refinement), a training-free two-stage framework. Its Adaptive Stage applies pivoted QR to construct an over-budget feature-coverage candidate pool, while its Progressive Stage uses evolving text-to-visual attention at selected decoder layers to prune a nested survivor set under a target layer-average token budget. Extensive experiments across seven LVLMs spanning multiple architectures and 18 image and video benchmarks demonstrate the effectiveness of STAR-Pro under aggressive pruning. On LLaVA-Video-7B, STAR-Pro reduces visual tokens by 90.5%, retains 92.7% of baseline performance, and achieves a $2.24\times$ measured inference speedup. Code is available at https://github.com/EasonAI-5589/starpro.

cs.CV↗

arXiv2026

DE-Venus: A Data-Efficient RLVR Framework for Large Language Models

Reinforcement learning with verifiable rewards (RLVR) improves large language model reasoning, but its practical scaling is constrained by expensive on-policy rollouts and the cost of obtaining reliable targets at scale. Existing methods address sample selection, incomplete supervision, or noisy labels separately, often entangling supervision logic with distributed training and hindering controlled comparison and reuse. We present DE-Venus, a unified framework for data-efficient RLVR that treats supervision as evolving state across data preparation and policy optimization. It organizes this lifecycle into three modules: Active Data Selection allocates training and annotation budgets; Weak Supervision Construction derives learning signals from unlabeled examples; and Training-Time Supervision Refinement filters or corrects unreliable supervision. DE-Venus supports seven representative methods and a data-selection pipeline by expressing method-specific decisions as offline dataset transitions or online transformations of targets, rewards, batches, and advantages while preserving verl's distributed execution contracts. Across public benchmarks and three business scenarios, separate configurations preserve or improve model quality with only 10% of labels or as little as 13% of relevant data; selected business configurations also reduce observed convergence steps by 63%--75%. DE-Venus thus reduces annotation and training costs without sacrificing scalable RL execution.

cs.LG↗

arXiv2026

Background Intensity Estimation for Cassini-ISS Image Using Deep Learning-Based Diffusion Model

Accurate background intensity estimation is crucial for precise astrometric measurements in astronomical imaging, particularly in complex scenarios such as those encountered in Cassini Imaging Science Subsystem (ISS) observations of Saturn's ring system. Traditional methods, like polynomial fitting and statistical method, often fail in non-uniform conditions, such as those caused by Saturn's rings or scattered light, due to mismatched assumptions and reliance on prior knowledge. This results in biased estimates and poor generalizability. We propose a deep learning framework based on Denoising Diffusion Probabilistic Model (DDPM) to address these challenges. By learning noise patterns and iteratively reconstructing backgrounds, DDPM improve background intensity estimation in ring-gap regions of ISS images by up to 62% relative to polynomial fitting. Additionally, When applied to centroiding of unresolved satellites in ring-gap, DDPM-based background estimation enhances positional precision by about 14% in the line direction and 15% in the sample direction. The framework autonomously captures spatial correlations, requires no manual parameter tuning, and generalizes across diverse backgrounds. These characteristics make it a promising, assumption-light solution for background estimation tasks in astrometry, photometry, and source detection, with potential applications to exoplanet transit imaging, deep-field surveys, and future missions.

astro-ph.IM↗

arXiv2026

Hermitian manifolds with nonpositive holomorphic sectional curvature

We study compact Kähler manifolds admitting Hermitian metrics with nonpositive holomorphic sectional curvature. We prove that the canonical bundle of such a manifold is nef, removing the pluriclosed assumption from the corresponding nefness result of Broder-Stanfield \cite{BroderStanfield}. In complex dimension two, we further show that negative holomorphic sectional curvature implies the ampleness of the canonical bundle. In complex dimension three, the same conclusion holds if the Hermitian metric is additionally assumed to be balanced. We also prove that vanishing holomorphic sectional curvature forces the first Chern class to vanish.

math.DG↗

arXiv2026

Automated Shape-Model-Based Astrometry of Phobos from Mars Express SRC Images

High-resolution spacecraft images provide important astrometric constraints for orbit refinement, but measurements of resolved bodies are often limited by labor-intensive control-point selection and the difficulty of achieving consistent reductions over large image archives. We present an automated shape-model-based astrometric pipeline for Phobos and apply it to Mars Express Super Resolution Channel (SRC) images. For each exposure, a synthetic image is rendered from a high-resolution 3D shape model under the nominal spacecraft-target-Sun geometry. Feature correspondences between the observed and synthetic images are established using SuperPoint and SuperGlue, followed by RANSAC filtering. The matched synthetic-image keypoints are then associated with surface points through ray-shape intersection. The geometric adjustment fixes the adopted body orientation, spacecraft state, and corrected camera pointing and estimates only two effective plane-of-sky position offsets using the exact perspective-projection model. These offsets are used to derive the center-of-figure position of Phobos. We first test the method on an image set previously analysed with a control-point approach and obtain comparable astrometric performance. We then extend the analysis to a larger SRC dataset spanning 2007-2025 and obtain 1113 successful measurements. Relative to the JPL MAR099 ephemeris, the resulting observed-minus-computed residuals have mean values of 0.186 km in $α\times cos(δ)$ and 0.053 km in $δ$, with corresponding standard deviations of 0.609 km and 0.583 km. These results demonstrate that the proposed pipeline provides a practical approach to large-scale, homogeneous astrometric reduction of archival spacecraft images of Phobos, with potential application to other resolved bodies.

astro-ph.IM↗

arXiv2026

Precise Top-Layer Fabric Segmentation for Fabric Destacking with Edge- and Shape-Aware Deep Networks

Fabric destacking requires precise segmentation of the topmost fabric layer, a task complicated by subtle fabric boundaries and high visual similarity between fabric layers. Existing semantic and edge-based segmentation approaches often struggle with these complexities, limiting the performance of robotic manipulation for different tasks. In this work, a novel segmentation training architecture tailored for top-layer fabric segmentation in stacked fabrics is proposed. The method extends the classical encoder-decoder framework by introducing two specialized branches - an edge-aware branch and a shape-aware branch - that are used to supervise the backbone network for better tuning. The edge-aware branch enhances boundary delineation, while the shape-aware branch guides the network to capture and align the overall fabric shape with reference masks derived from Computer Aided Design (CAD) models. Experiments on a real-world fabric dataset demonstrate that the training approach outperforms established baselines, verifying the effectiveness of the multi-branch design through both quantitative results and ablation studies.

cs.CV↗

arXiv2026

Robotic Fabric Alignment System for Sewing Using Global Local Weighted ICP

Accurate fabric alignment is a critical step that must be performed before sewing. This paper presents a novel automated fabric alignment system. The system estimates the poses of top and bottom fabric panels, lying flat and wrinkle-free in arbitrary positions, using a new Global Local Weighted Iterative Closest Point (GLW-ICP) method. The system then manipulates the top panel to achieve precise alignment at both edges and sewing lines. Unlike conventional approaches, GLW-ICP robustly aligns both global edges and local sewing lines by globally aligning fabric edge points and locally aligning sewing line points to their corresponding CAD model points, while removing unmatched points in occluded regions. Real-world experiments with various fabric shapes show that the system consistently achieves millimeter-level alignment accuracy under both occlusion and non-occlusion conditions, demonstrating its effectiveness and suitability for automated fabric alignment in practical scenarios.

cs.RO↗