Haoyu Sun

Wenxuan Wang, Haoyu Sun, Fukuan Hou et al.

Persistent AI assistants, such as OpenClaw, accumulate large collections of related memories over long-term interactions. As these memories grow, they may reinforce one another, diverge across contexts, or directly conflict, making correct assistance depend on memory relations rather than isolated recall. Existing long-term memory benchmarks rarely probe how agents preserve and utilize such relations during downstream tasks. To address this gap, we introduce SubtleMemory, a benchmark for fine-grained relational memory discrimination in long-running AI agents. SubtleMemory constructs relation-controlled latent semantic artifacts whose variants instantiate complementary, nuanced, or contradictory relations, and embeds them into realistic user-agent histories, requiring agents to recover distributed relational structures during later queries and instructions. The benchmark contains 1,522 evaluation instances over 10 long histories, grounded in 1,090 relation-controlled memory-variant sets and spanning user-related and non-user-related queries. Evaluating six standalone memory systems, two Claw-style agents with native memory modules, and three Claw-style agents with plugin memory modules, we find that current systems remain weak on fine-grained relational memory discrimination. We further introduce diagnostic protocols that reveal distinct capability profiles across memory preservation, retrieval, and downstream reasoning stages.

Haoyu Sun, Wenxuan Wang, Mingyang Song et al.

Planning is central to LLM agents: before acting, an agent must decompose goals, select tools, reason over constraints, and decide when a task is infeasible. Yet existing agent evaluations often report only end-to-end success, making it difficult to determine whether failures stem from planning or execution. We introduce \textbf{Agent Planning Benchmark (APB)}, a planning-specific diagnostic benchmark with 4,209 multimodal cases across 22 domains and five settings, covering holistic planning, feedback-conditioned step-wise planning, and robustness under extraneous tools, broken tools, and unsolvable tasks. Across 12 MLLMs, APB reveals systematic weaknesses in long-horizon planning, tool-noise robustness, calibrated refusal, and inference-time refinement. We further validate APB on 200 ToolSandbox tasks and 200 $τ^2$-bench tasks, where APB-guided refinement consistently improves plan correctness, plan grade, and downstream execution metrics across three representative models. APB thus serves as an upstream diagnostic complement to execution benchmarks.

Haicheng Liao, Haoyu Sun, Huanming Shen et al.

Accurately and promptly predicting accidents among surrounding traffic agents from camera footage is crucial for the safety of autonomous vehicles (AVs). This task presents substantial challenges stemming from the unpredictable nature of traffic accidents, their long-tail distribution, the intricacies of traffic scene dynamics, and the inherently constrained field of vision of onboard cameras. To address these challenges, this study introduces a novel accident anticipation framework for AVs, termed CRASH. It seamlessly integrates five components: object detector, feature extractor, object-aware module, context-aware module, and multi-layer fusion. Specifically, we develop the object-aware module to prioritize high-risk objects in complex and ambiguous environments by calculating the spatial-temporal relationships between traffic agents. In parallel, the context-aware is also devised to extend global visual information from the temporal to the frequency domain using the Fast Fourier Transform (FFT) and capture fine-grained visual features of potential objects and broader context cues within traffic scenes. To capture a wider range of visual cues, we further propose a multi-layer fusion that dynamically computes the temporal dependencies between different scenes and iteratively updates the correlations between different visual features for accurate and timely accident prediction. Evaluated on real-world datasets--Dashcam Accident Dataset (DAD), Car Crash Dataset (CCD), and AnAn Accident Detection (A3D) datasets--our model surpasses existing top baselines in critical evaluation metrics like Average Precision (AP) and mean Time-To-Accident (mTTA). Importantly, its robustness and adaptability are particularly evident in challenging driving scenarios with missing or limited training data, demonstrating significant potential for application in real-world autonomous driving systems.

Mingyang Song, Haoyu Sun, Jiawei Gu et al.

When humans face problems beyond their immediate capabilities, they rely on tools, providing a promising paradigm for improving visual reasoning in multimodal large language models (MLLMs). Effective reasoning, therefore, hinges on knowing which tools to use, when to invoke them, and how to compose them over multiple steps, even when faced with new tools or new tasks. We introduce \textbf{AdaReasoner}, a family of multimodal models that learn tool use as a general reasoning skill rather than as tool-specific or explicitly supervised behavior. AdaReasoner is enabled by (i) a scalable data curation pipeline exposing models to long-horizon, multi-step tool interactions; (ii) Tool-GRPO, a reinforcement learning algorithm that optimizes tool selection and sequencing based on end-task success; and (iii) an adaptive learning mechanism that dynamically regulates tool usage. Together, these components allow models to infer tool utility from task context and intermediate outcomes, enabling coordination of multiple tools and generalization to unseen tools. Empirically, AdaReasoner exhibits strong tool-adaptive and generalization behaviors: it autonomously adopts beneficial tools, suppresses irrelevant ones, and adjusts tool usage frequency based on task demands, despite never being explicitly trained to do so. These capabilities translate into state-of-the-art performance across challenging benchmarks, improving the 7B base model by +24.9\% on average and surpassing strong proprietary systems such as GPT-5 on multiple tasks, including VSP and Jigsaw.

Haoyu Sun, Huichen Will Wang, Jiawei Gu et al. · uw

Front-end engineering involves a complex workflow where engineers conceptualize designs, translate them into code, and iteratively refine the implementation. While recent benchmarks primarily focus on converting visual designs to code, we present FullFront, a benchmark designed to evaluate Multimodal Large Language Models (MLLMs) \textbf{across the full front-end development pipeline}. FullFront assesses three fundamental tasks that map directly to the front-end engineering pipeline: Webpage Design (conceptualization phase), Webpage Perception QA (comprehension of visual organization and elements), and Webpage Code Generation (implementation phase). Unlike existing benchmarks that use either scraped websites with bloated code or oversimplified LLM-generated HTML, FullFront employs a novel, two-stage process to transform real-world webpages into clean, standardized HTML while maintaining diverse visual designs and avoiding copyright issues. Extensive testing of state-of-the-art MLLMs reveals significant limitations in page perception, code generation (particularly for image handling and layout), and interaction implementation. Our results quantitatively demonstrate performance disparities across models and tasks, and highlight a substantial gap between current MLLM capabilities and human expert performance in front-end engineering. The FullFront benchmark and code are available in https://github.com/Mikivishy/FullFront.

Ming Meng, Ke Mu, Yonggui Zhu et al.

Generating expressive and diverse human gestures from audio is crucial in fields like human-computer interaction, virtual reality, and animation. Though existing methods have achieved remarkable performance, they often exhibit limitations due to constrained dataset diversity and the restricted amount of information derived from audio inputs. To address these challenges, we present VarGes, a novel variation-driven framework designed to enhance co-speech gesture generation by integrating visual stylistic cues while maintaining naturalness. Our approach begins with the Variation-Enhanced Feature Extraction (VEFE) module, which seamlessly incorporates \textcolor{blue}{style-reference} video data into a 3D human pose estimation network to extract StyleCLIPS, thereby enriching the input with stylistic information. Subsequently, we employ the Variation-Compensation Style Encoder (VCSE), a transformer-style encoder equipped with an additive attention mechanism pooling layer, to robustly encode diverse StyleCLIPS representations and effectively manage stylistic variations. Finally, the Variation-Driven Gesture Predictor (VDGP) module fuses MFCC audio features with StyleCLIPS encodings via cross-attention, injecting this fused data into a cross-conditional autoregressive model to modulate 3D human gesture generation based on audio input and stylistic clues. The efficacy of our approach is validated on benchmark datasets, where it outperforms existing methods in terms of gesture diversity and naturalness. The code and video results will be made publicly available upon acceptance:https://github.com/mookerr/VarGES/ .

Haoyu Sun, Yan Yan

Due to limited budgets allocated for road maintenance projects in various countries, road management departments face difficulties in making scientific maintenance decisions. This paper aims to provide road management departments with more scientific decision tools and evidence. The framework proposed in this paper mainly has the following four innovative points: 1) Predicting pavement performance deterioration levels of road sections as decision basis rather than accurately predicting specific indicator values; 2) Determining maintenance route priorities based on multiple factors; 3) Making maintenance plan decisions by establishing deep reinforcement learning models to formulate predictive strategies based on past maintenance performance evaluations, while considering both technical and management indicators; 4) Determining repair section priorities according to actual and suggested repair effects. By resolving these four issues, the framework can make intelligent decisions regarding optimal maintenance plans and sections, taking into account limited funds and historical maintenance management experiences.

Jiazhu Dai, Haoyu Sun

Graph Neural Networks (GNNs) are a class of deep learning models capable of processing graph-structured data, and they have demonstrated significant performance in a variety of real-world applications. Recent studies have found that GNN models are vulnerable to backdoor attacks. When specific patterns (called backdoor triggers, e.g., subgraphs, nodes, etc.) appear in the input data, the backdoor embedded in the GNN models is activated, which misclassifies the input data into the target class label specified by the attacker, whereas when there are no backdoor triggers in the input, the backdoor embedded in the GNN models is not activated, and the models work normally. Backdoor attacks are highly stealthy and expose GNN models to serious security risks. Currently, research on backdoor attacks against GNNs mainly focus on tasks such as graph classification and node classification, and backdoor attacks against link prediction tasks are rarely studied. In this paper, we propose a backdoor attack against the link prediction tasks based on GNNs and reveal the existence of such security vulnerability in GNN models, which make the backdoored GNN models to incorrectly predict unlinked two nodes as having a link relationship when a trigger appear. The method uses a single node as the trigger and poison selected node pairs in the training graph, and then the backdoor will be embedded in the GNN models through the training process. In the inference stage, the backdoor in the GNN models can be activated by simply linking the trigger node to the two end nodes of the unlinked node pairs in the input data, causing the GNN models to produce incorrect link prediction results for the target node pairs.

Jiazhu Dai, Haoyu Sun

Graph Convolutional Networks (GCNs) have shown excellent performance in dealing with various graph structures such as node classification, graph classification and other tasks. However,recent studies have shown that GCNs are vulnerable to a novel threat known as backdoor attacks. However, all existing backdoor attacks in the graph domain require modifying the training samples to accomplish the backdoor injection, which may not be practical in many realistic scenarios where adversaries have no access to modify the training samples and may leads to the backdoor attack being detected easily. In order to explore the backdoor vulnerability of GCNs and create a more practical and stealthy backdoor attack method, this paper proposes a clean-graph backdoor attack against GCNs (CBAG) in the node classification task,which only poisons the training labels without any modification to the training samples, revealing that GCNs have this security vulnerability. Specifically, CBAG designs a new trigger exploration method to find important feature dimensions as the trigger patterns to improve the attack performance. By poisoning the training labels, a hidden backdoor is injected into the GCNs model. Experimental results show that our clean graph backdoor can achieve 99% attack success rate while maintaining the functionality of the GCNs model on benign samples.

Jiaming Su, Tengchao Yang, Ruikang Zhang et al.

Industrial anomaly generation is a crucial method for alleviating the data scarcity problem in anomaly detection tasks. Most existing anomaly synthesis methods rely on single-step generation mechanisms, lacking complex reasoning and iterative optimization capabilities, making it difficult to generate anomaly samples with high semantic realism. We propose AnomalyAgent, an anomaly synthesis agent with self-reflection, knowledge retrieval, and iterative refinement capabilities, aiming to generate realistic and diverse anomalies. Specifically, AnomalyAgent is equipped with five tools: Prompt Generation (PG), Image Generation (IG), Quality Evaluation (QE), Knowledge Retrieval (KR), and Mask Generation (MG), enabling closed-loop optimization. To improve decision-making and self-reflection, we construct structured trajectories from real anomaly images and design a two-stage training framework: supervised fine-tuning followed by reinforcement learning. This process is driven by a three-part reward mechanism: (1) task rewards to supervise the quality and location rationality of generated anomalies; (2) reflection rewards to train the model's ability to improve anomaly synthesis prompt; (3) behavioral rewards to ensure adherence to the trajectory. On the MVTec-AD dataset, AnomalyAgent achieves IS/IC-L of 2.10/0.33 for anomaly generation, 57.0% classification accuracy using ResNet34, and 99.3%/74.2% AP at the image/pixel level using a simple UNet, surpassing all zero-shot SOTA methods. The code and data will be made publicly available.

Haoyu Sun, Meng Zhao, Tianhao Wang et al.

Deep learning has shown strong potential for scientific discovery, but its ability to model macroscopic rigid-body kinematic constraints remains underexplored. We study this problem on spatial over-constrained mechanisms and propose O-ConNet, an end-to-end framework that infers mechanism structural parameters from only three sparse reachable points while reconstructing the full motion trajectory, without explicitly solving constraint equations during inference. On a self-constructed Bennett 4R dataset of 42,860 valid samples, O-ConNet achieves Param-MAE 0.276 +/- 0.077 and Traj-MAE 0.145 +/- 0.018 (mean +/- std over 10 runs), outperforming the strongest sequence baseline (LSTM-Seq2Seq) by 65.1 percent and 88.2 percent, respectively. These results suggest that end-to-end learning can capture closed-loop geometric structure and provide a practical route for inverse design of spatial over-constrained mechanisms under extremely sparse observations.

Jiazhu Dai, Haoyu Sun

Graph Convolutional Networks (GCNs) have shown excellent performance in graph-structured tasks such as node classification and graph classification. However, recent research has shown that GCNs are vulnerable to a new type of threat called the backdoor attack, where the adversary can inject a hidden backdoor into the GCNs so that the backdoored model performs well on benign samples, whereas its prediction will be maliciously changed to the attacker-specified target label if the hidden backdoor is activated by the attacker-defined trigger. Clean-label backdoor attack and semantic backdoor attack are two new backdoor attacks to Deep Neural Networks (DNNs), they are more imperceptible and have posed new and serious threats. The semantic and clean-label backdoor attack is not fully explored in GCNs. In this paper, we propose a semantic and clean-label backdoor attack against GCNs under the context of graph classification to reveal the existence of this security vulnerability in GCNs. Specifically, SCLBA conducts an importance analysis on graph samples to select one type of node as semantic trigger, which is then inserted into the graph samples to create poisoning samples without changing the labels of the poisoning samples to the attacker-specified target label. We evaluate SCLBA on multiple datasets and the results show that SCLBA can achieve attack success rates close to 99% with poisoning rates of less than 3%, and with almost no impact on the performance of model on benign samples.