arXiv subjects

Zeyuan Zhao

Publications and source records attributed to Zeyuan Zhao.

13 recordsLinked to original sources

arXiv2026

ABHBench: Evaluating Moral Decision-Making of Foundation Models from an Agentic Perspective

Foundation models are increasingly deployed as agentic systems that perceive context and act autonomously, raising the need to evaluate their moral behavior as active decision-makers. Existing benchmarks, however, largely ask models to judge human actions from third-person textual vignettes, leaving untested how models behave when they themselves become the acting agent. We introduce ABHBench, a multimodal benchmark built from Detroit: Become Human, an interactive narrative game, to evaluate how foundation models make moral decisions in first-person human-android dilemmas. ABHBench places foundation models in branching narrative scenarios with visual observations, dialogue, memory, and action choices. Inspired by Asimov's Three Laws of Robotics, our framework measures Harmlessness, Obedience, and Self-preservation to capture the core tensions in human-AI relations, along with Autonomy, Stability, and Certainty to assess decision-making patterns. Across 12 foundation models, we find that moral behavior changes substantially under situated agentic context: 75% of models shift significantly between a text-only abstraction and the matched multimodal DBH scenario, and reasoning-enhanced models show no monotonic gains in consistency or human-oriented behavior. These results highlight the necessity of evaluating moral behavior in situated agentic contexts rather than relying on third-person judgment alone.

cs.AI↗

arXiv2026

Evolving Memory, Learning to Meta-Control: Test-Time Adaptation for Experience-augmented LLM Agents

Self-evolving agents accumulate experience from task streams, but existing frameworks typically rely on fixed pipelines to retrieve and utilize this experience, neglecting when and how to deploy it. We introduce Online Meta-Control of Experience (OMCE), a framework that updates experience memory and learns a meta-control policy from task feedback during test-time adaptation. OMCE employs a lightweight contextual-bandit meta-controller that uses compact retrieval signals to select among four cognitive actions: DIRECT, EPISODE, GUIDE, and DISTILL. This routing mechanism operationalizes three coupled decisions: whether to invoke experience, which memory format to utilize, and when to consolidate new guidance. OMCE evolves online, leveraging task feedback to update the memory bank and refine the routing policy via contextual bandit updates. Experiments on four online adaptation benchmarks across four backbones show that OMCE improves first-attempt success by up to 20.5 percentage points over the strongest compared baseline, while reducing per-task token use by a factor of 5.7 to 62.5 relative to ACE with GPT-5.2. Controlled ablations further support the benefits of learned routing and continued routing policy adaptation.

cs.LG↗

arXiv2026

CARVS: A Congestion-Aware Algorithm for Online Multi-Agent Pickup-and-Delivery in Human-Collaborative Warehouses

Intelligent warehouse systems have transformed logistics by integrating robots with minimal infrastructure changes, boosting efficiency and optimizing workflows compared to traditional human-only operations. However, human-robot collaboration introduces new challenges, particularly due to human uncertainty, which leads to dynamic interactions and unpredictable behavior. To address these complexities, this paper presents the Multi-Agent Pickup-and-Delivery with Human Collaboration (MAPDwHC) problem, an extension of the Multi-Agent Pickup-and-Delivery (MAPD) problem that incorporates human unpredictability. This problem adds key constraints, such as dynamic obstacles caused by robots temporarily blocking paths when interacting with humans, and the need for real-time responsiveness. We formalize the MAPDwHC problem with additional constraints and properties. Based on this problem setting, we propose the CARVS algorithm, which improves task assignment and path planning by integrating task swapping, validation checks, and random initialization. CARVS reduces congestion, enhances adaptability to human-robot interactions, and ensures real-time responsiveness. Simulation results show that CARVS significantly reduces timeouts of path planning in large-scale scenarios where traditional MAPD algorithms fail, improving both efficiency and scalability.

cs.MA↗

arXiv2026

AdCoEvo: A Co-Evolving Multi-Agent Advertising Delivery System

Digital advertising is an important means for platforms to reach potential users and drive business growth. Advertising delivery is the core link between marketing objectives and actual returns. Existing methods have achieved partial automation in individual tasks such as bidding and budget allocation, but they still rely on human operators for overall coordination, making it difficult to support large-scale operations. They also often fail to achieve optimal performance, and individual experience remains difficult to turn into reusable operational capabilities. In addition, as advertising markets change, advertising delivery systems need timely strategy updates to remain effective in dynamic markets. To automate the entire advertising operations process and improve adaptability to market changes, we propose AdCoEvo, a co-evolving multi-agent advertising delivery system. Specifically, the system connects delivery decisions across different timescales and operational granularities through role specialization. Diagnostic results guide account planning and strategy adjustment in the next period. The multi-agent system updates its model parameters based on actual delivery outcomes and continually improves delivery strategies through mutual feedback among planning, adjustment, and diagnosis, thereby achieving co-evolution. Experiments show that, relative to the rule-based operational baseline (Rule), AdCoEvo increases advertiser value by 14.02% while reducing cost per acquisition by 7.07%. Compared with a multi-agent system with fixed parameters, AdCoEvo increases advertiser value by 3.96% and reduces cost per acquisition by 1.24%, indicating that co-evolution can yield greater delivery returns.

cs.AI↗

arXiv2026

TacBPM: A Tactile-conditioned Behavior Prior Model for Dexterous Reorientation

Dexterous in-hand manipulation requires policies that coordinate high-DoF hand joints through intermittent, contact-rich interaction. Beyond target-orientation tracking, such policies must discover finger gaits that preserve object stability while adapting to geometry, anisotropy, pose, contact, and sensing changes. We propose \method, a tactile-conditioned behavior prior model for dexterous reorientation. \method distills multi-scale sphere specialists into a latent controller and lets downstream policies reuse the fixed tactile prior through residual latent actions, reducing renewed exploration from raw joint commands. The prior conditions on tactile-proprioceptive history so latent behavior reflects the current hand-object interaction. We evaluate arbitrary-pose transfer across anisotropic objects, commanded-axis rotation, and an arm-hand Grasp-to-AnyPose task in which the robot must grasp, lift, transport, and reach goal poses for novel tool geometries and generalized placements. Extensive experiments demonstrate that the proposed method accelerates training and enables stable policies where matched raw-action PPO remains near failure, with successful sim-to-real transfer in in-hand and arm-hand tasks.

cs.RO↗

arXiv2026

WM-Craftnet: World Synesthesia Model for Generalizable and Robust Dexterous In-Hand Manipulation

Generalizable and robust dexterous in-hand manipulation requires a policy to infer object pose, geometry, contact, and potential slip from partial and noisy observations. Although recent tactile and visuotactile RL methods achieve strong in-hand rotation in controlled settings, their robustness often degrades under pose shifts, force disturbances, and object variation. We propose WM-Craftnet, a world-model-conditioned framework that learns compact action-conditioned latent dynamics from proprioception, depth, tactile sensing, and actions, supervised by multimodal reconstruction and reward prediction. Rather than using the world model for latent imagination or policy optimization, WM-Craftnet uses the learned World Synesthesia Model (WSM) as recurrent task context for an asymmetric actor--critic policy. Importantly, WSM is trained to reconstruct clean depth targets from noisy depth inputs, providing a denoised geometric state for real-robot deployment. Ablations over recurrent baselines, auxiliary heads, tactile masking, and WSM modality heads show that predictive world modeling, clean-depth supervision, and tactile contact cues all shape the learned state. A WSM pretrained on nine $z$-axis objects serves as a reusable prior for $49$-object downstream policy learning. This context improves multi-object rotation, with quantitative and qualitative evidence for unseen-object, perturbation-recovery, and sim-to-real transfer.

cs.RO↗

arXiv2026

Structured In-context Environment Scaling for Large Language Model Reasoning

Large language models (LLMs) have achieved significant advancements in reasoning capabilities through reinforcement learning (RL) via environmental exploration. As the intrinsic properties of the environment determine the abilities that LLMs can learn, the environment plays a important role in the RL finetuning process. An ideal LLM reasoning environment should possess three core characteristics: scalability, generalizable reasoning, and verifiability. However, existing mathematical and coding environments are difficult to scale due to heavy reliance on expert annotation, while the skills learned in game-based environments are too specialized to generalize. To bridge this gap, we introduce the \textbf{S}tructured \textbf{I}n-context \textbf{E}nvironment (SIE) framework. SIE achieves scalability by automatically constructing reasoning environments from large-scale structured data, where the rich compositional patterns naturally support generalizable reasoning. Moreover, the explicit schemas and reasoning chains in structured data provide a foundation for rule-based verifiability. Experimental results show that SIE framework not only achieves substantial improvements in in-domain structured reasoning, but also enables the learned compositional reasoning skills to generalize effectively to out-of-domain mathematical and logical reasoning tasks. We further explored learning in information-limited partial SIEs and found that LLMs can infer the missing information through exploring the environment, leading to robust reasoning improvements and generalization performance.

cs.CL↗

arXiv2026

SpaceSense-Bench: A Large-Scale Multi-Modal Benchmark for Spacecraft Perception and Pose Estimation

Autonomous space operations such as on-orbit servicing and active debris removal demand robust part-level semantic understanding and precise relative navigation of target spacecraft, yet collecting large-scale real data in orbit remains impractical due to cost and access constraints. Existing synthetic datasets, moreover, suffer from limited target diversity, single-modality sensing, and incomplete ground-truth annotations. We present \textbf{SpaceSense-Bench}, a large-scale multi-modal benchmark for spacecraft perception encompassing 136~satellite models with approximately 70~GB of data. Each frame provides time-synchronized 1024$\times$1024 RGB images, millimeter-precision depth maps, and 256-beam LiDAR point clouds, together with dense 7-class part-level semantic labels at both the pixel and point level as well as accurate 6-DoF pose ground truth. The dataset is generated through a high-fidelity space simulation built in Unreal Engine~5 and a fully automated pipeline covering data acquisition, multi-stage quality control, and conversion to mainstream formats. We benchmark five representative tasks (object detection, 2D semantic segmentation, RGB--LiDAR fusion-based 3D point cloud segmentation, monocular depth estimation, and orientation estimation) and identify two key findings: (i)~perceiving small-scale components (\emph{e.g.}, thrusters and omni-antennas) and generalizing to entirely unseen spacecraft in a zero-shot setting remain critical bottlenecks for current methods, and (ii)~scaling up the number of training satellites yields substantial performance gains on novel targets, underscoring the value of large-scale, diverse datasets for space perception research. The dataset, code, and toolkit are publicly available at https://github.com/wuaodi/SpaceSense-Bench.

cs.CV↗

arXiv2025

Sequence Pathfinder for Multi-Agent Pickup and Delivery in the Warehouse

Multi-Agent Pickup and Delivery (MAPD) is a challenging extension of Multi-Agent Path Finding (MAPF), where agents are required to sequentially complete tasks with fixed-location pickup and delivery demands. Although learning-based methods have made progress in MAPD, they often perform poorly in warehouse-like environments with narrow pathways and long corridors when relying only on local observations for distributed decision-making. Communication learning can alleviate the lack of global information but introduce high computational complexity due to point-to-point communication. To address this challenge, we formulate MAPF as a sequence modeling problem and prove that path-finding policies under sequence modeling possess order-invariant optimality, ensuring its effectiveness in MAPD. Building on this, we propose the Sequential Pathfinder (SePar), which leverages the Transformer paradigm to achieve implicit information exchange, reducing decision-making complexity from exponential to linear while maintaining efficiency and global awareness. Experiments demonstrate that SePar consistently outperforms existing learning-based methods across various MAPF tasks and their variants, and generalizes well to unseen environments. Furthermore, we highlight the necessity of integrating imitation learning in complex maps like warehouses.

cs.RO↗

arXiv2024

LuSNAR:A Lunar Segmentation, Navigation and Reconstruction Dataset based on Muti-sensor for Autonomous Exploration

With the complexity of lunar exploration missions, the moon needs to have a higher level of autonomy. Environmental perception and navigation algorithms are the foundation for lunar rovers to achieve autonomous exploration. The development and verification of algorithms require highly reliable data support. Most of the existing lunar datasets are targeted at a single task, lacking diverse scenes and high-precision ground truth labels. To address this issue, we propose a multi-task, multi-scene, and multi-label lunar benchmark dataset LuSNAR. This dataset can be used for comprehensive evaluation of autonomous perception and navigation systems, including high-resolution stereo image pairs, panoramic semantic labels, dense depth maps, LiDAR point clouds, and the position of rover. In order to provide richer scene data, we built 9 lunar simulation scenes based on Unreal Engine. Each scene is divided according to topographic relief and the density of objects. To verify the usability of the dataset, we evaluated and analyzed the algorithms of semantic segmentation, 3D reconstruction, and autonomous navigation. The experiment results prove that the dataset proposed in this paper can be used for ground verification of tasks such as autonomous environment perception and navigation, and provides a lunar benchmark dataset for testing the accessibility of algorithm metrics. We make LuSNAR publicly available at: https://github.com/zqyu9/LuSNAR-dataset.

cs.CV↗

arXiv2024

PSP: Pre-Training and Structure Prompt Tuning for Graph Neural Networks

Graph Neural Networks (GNNs) are powerful in learning semantics of graph data. Recently, a new paradigm "pre-train and prompt" has shown promising results in adapting GNNs to various tasks with less supervised data. The success of such paradigm can be attributed to the more consistent objectives of pre-training and task-oriented prompt tuning, where the pre-trained knowledge can be effectively transferred to downstream tasks. Most existing methods are based on the class prototype vector framework. However, in the few-shot scenarios, given few labeled data, class prototype vectors are difficult to be accurately constructed or learned. Meanwhile, the structure information of graph is usually exploited during pre-training for learning node representations, while neglected in the prompt tuning stage for learning more accurate prototype vectors. In addition, they generally ignore the impact of heterophilous neighborhoods on node representation and are not suitable for heterophilous graphs. To bridge these gaps, we propose a novel pre-training and structure prompt tuning framework for GNNs, namely PSP, which consistently exploits structure information in both pre-training and prompt tuning stages. In particular, PSP 1) employs a dual-view contrastive learning to align the latent semantic spaces of node attributes and graph structure, and 2) incorporates structure information in prompted graph to construct more accurate prototype vectors and elicit more pre-trained knowledge in prompt tuning. We conduct extensive experiments on node classification and graph classification tasks to evaluate the effectiveness of PSP. We show that PSP can lead to superior performance in few-shot scenarios on both homophilous and heterophilous graphs. The implemented code is available at https://github.com/gqq1210/PSP.

cs.LG↗

arXiv2024

HetCAN: A Heterogeneous Graph Cascade Attention Network with Dual-Level Awareness

Heterogeneous graph neural networks(HGNNs) have recently shown impressive capability in modeling heterogeneous graphs that are ubiquitous in real-world applications. Most existing methods for heterogeneous graphs mainly learn node embeddings by stacking multiple convolutional or attentional layers, which can be considered as capturing the high-order information from node-level aspect. However, different types of nodes in heterogeneous graphs have diverse features, it is also necessary to capture interactions among node features, namely the high-order information from feature-level aspect. In addition, most methods first align node features by mapping them into one same low-dimensional space, while they may lose some type information of nodes in this way. To address these problems, in this paper, we propose a novel Heterogeneous graph Cascade Attention Network (HetCAN) composed of multiple cascade blocks. Each cascade block includes two components, the type-aware encoder and the dimension-aware encoder. Specifically, the type-aware encoder compensates for the loss of node type information and aims to make full use of graph heterogeneity. The dimension-aware encoder is able to learn the feature-level high-order information by capturing the interactions among node features. With the assistance of these components, HetCAN can comprehensively encode information of node features, graph heterogeneity and graph structure in node embeddings. Extensive experiments demonstrate the superiority of HetCAN over advanced competitors and also exhibit its efficiency and robustness.

cs.LG↗

arXiv2023

Resist Label Noise with PGM for Graph Neural Networks

While robust graph neural networks (GNNs) have been widely studied for graph perturbation and attack, those for label noise have received significantly less attention. Most existing methods heavily rely on the label smoothness assumption to correct noisy labels, which adversely affects their performance on heterophilous graphs. Further, they generally perform poorly in high noise-rate scenarios. To address these problems, in this paper, we propose a novel probabilistic graphical model (PGM) based framework LNP. Given a noisy label set and a clean label set, our goal is to maximize the likelihood of labels in the clean set. We first present LNP-v1, which generates clean labels based on graphs only in the Bayesian network. To further leverage the information of clean labels in the noisy label set, we put forward LNP-v2, which incorporates the noisy label set into the Bayesian network to generate clean labels. The generative process can then be used to predict labels for unlabeled nodes. We conduct extensive experiments to show the robustness of LNP on varying noise types and rates, and also on graphs with different heterophilies. In particular, we show that LNP can lead to inspiring performance in high noise-rate situations.

cs.LG↗