Pushing the Limits of Safety: A Technical Report on the ATLAS Challenge 2025

Multimodal Large Language Models (MLLMs) have enabled transformative advancements across diverse applications but remain susceptible to safety threats, especially jailbreak attacks that induce harmful outputs. To systematically evaluate and improve their safety, we organized the Adversarial Testing & Large-model Alignment Safety Grand Challenge (ATLAS) 2025}. This technical report presents findings from the competition, which involved 86 teams testing MLLM vulnerabilities via adversarial image-text attacks in two phases: white-box and black-box evaluations. The competition results highlight ongoing challenges in securing MLLMs and provide valuable guidance for developing stronger defense mechanisms. The challenge establishes new benchmarks for MLLM safety evaluation and lays groundwork for advancing safer multimodal AI systems. The code and data for this challenge are openly available at https://github.com/NY1024/ATLAS_Challenge_2025.

USTBench: Benchmarking and Dissecting Spatiotemporal Reasoning of LLMs as Urban Agents

Large language models (LLMs) have shown emerging potential in spatiotemporal reasoning, making them promising candidates for building urban agents that support diverse urban downstream applications. Despite these benefits, existing studies primarily focus on evaluating urban LLM agent on outcome-level metrics (e.g., prediction accuracy, traffic efficiency), offering limited insight into their underlying reasoning processes. As a result, the strengths and limitations of urban LLM agents in spatiotemporal reasoning remain poorly understood. To this end, we introduce USTBench, the first benchmark to evaluate LLMs' spatiotemporal reasoning abilities as urban agents across four decomposed dimensions: spatiotemporal understanding, forecasting, planning, and reflection with feedback. Specifically, USTBench supports five diverse urban decision-making and four spatiotemporal prediction tasks, all running within our constructed interactive city environment UAgentEnv. The benchmark includes 62,466 structured QA pairs for process-level evaluation and standardized end-to-end task assessments, enabling fine-grained diagnostics and broad task-level comparison across diverse urban scenarios. Through extensive evaluation of thirteen leading LLMs, we reveal that although LLMs show promising potential across various urban downstream tasks, they still struggle in long-horizon planning and reflective adaptation in dynamic urban contexts. Notably, recent advanced reasoning models (e.g., DeepSeek-R1) trained on general logic or mathematical problems do not consistently outperform non-reasoning LLMs. This discrepancy highlights the need for domain-specialized adaptation methods to enhance urban spatiotemporal reasoning. Overall, USTBench provides a foundation to build more adaptive and effective LLM-based urban agents and broad smart city applications.

CoLLMLight: Cooperative Large Language Model Agents for Network-Wide Traffic Signal Control

Traffic Signal Control (TSC) plays a critical role in urban traffic management by optimizing traffic flow and mitigating congestion. While Large Language Models (LLMs) have recently emerged as promising tools for TSC due to their exceptional problem-solving and generalization capabilities, existing approaches fail to address the essential need for inter-agent coordination, limiting their effectiveness in achieving network-wide optimization. To bridge this gap, we propose CoLLMLight, a cooperative LLM agent framework for TSC. Specifically, we first construct a structured spatiotemporal graph to capture real-time traffic dynamics and spatial relationships among neighboring intersections, enabling the LLM to reason about complex traffic interactions. Moreover, we introduce a complexity-aware reasoning mechanism that dynamically adapts reasoning depth based on real-time traffic conditions, ensuring optimal computational efficiency without sacrificing decision quality. Besides, we propose a fine-tuning strategy that leverages iterative simulation-driven data collection and environmental feedback to build a lightweight LLM tailored for cooperative TSC. Extensive experiments on both synthetic and real-world datasets demonstrate that CoLLMLight outperforms state-of-the-art methods in diverse traffic scenarios, showcasing its effectiveness, scalability, and robustness.

Large Language Models as Traffic Signal Control Agents: Capacity and Opportunity

Traffic signal control is crucial for optimizing the efficiency of road network by regulating traffic light phases. Existing research predominantly focuses on heuristic or reinforcement learning (RL)-based methods, which often lack transferability across diverse traffic scenarios and suffer from poor interpretability. This paper introduces a novel approach, LLMLight, utilizing large language models (LLMs) for traffic signal control tasks. By leveraging LLMs' impressive generalization and zero-shot reasoning capabilities, LLMLight executes a human-like decision-making process for efficient traffic management. Specifically, the framework begins by composing task descriptions, current traffic conditions, and prior knowledge into a prompt. Subsequently, we utilize LLM's chain-of-thought (CoT) reasoning ability to identify the next traffic signal phase, ensuring optimal efficiency in the road network. LLMLight achieves state-of-the-art (SOTA) or competitive results across five real-world traffic datasets. Notably, LLMLight showcases remarkable generalization, interpretability, and zero-shot reasoning abilities, even without any training for transportation management tasks. Our project is available at https://github.com/usail-hkust/LLMTSCS.

A Preference-aware Meta-optimization Framework for Personalized Vehicle Energy Consumption Estimation

Vehicle Energy Consumption (VEC) estimation aims to predict the total energy required for a given trip before it starts, which is of great importance to trip planning and transportation sustainability. Existing approaches mainly focus on extracting statistically significant factors from typical trips to improve the VEC estimation. However, the energy consumption of each vehicle may diverge widely due to the personalized driving behavior under varying travel contexts. To this end, this paper proposes a preference-aware meta-optimization framework Meta-Pec for personalized vehicle energy consumption estimation. Specifically, we first propose a spatiotemporal behavior learning module to capture the latent driver preference hidden in historical trips. Moreover, based on the memorization of driver preference, we devise a selection-based driving behavior prediction module to infer driver-specific driving patterns on a given route, which provides additional basis and supervision signals for VEC estimation. Besides, a driver-specific meta-optimization scheme is proposed to enable fast model adaption by learning and sharing transferable knowledge globally. Extensive experiments on two real-world datasets show the superiority of our proposed framework against ten numerical and data-driven machine learning baselines. The source code is available at https://github.com/usail-hkust/Meta-Pec.

Bkd-FedGNN: A Benchmark for Classification Backdoor Attacks on Federated Graph Neural Network

Federated Graph Neural Network (FedGNN) has recently emerged as a rapidly growing research topic, as it integrates the strengths of graph neural networks and federated learning to enable advanced machine learning applications without direct access to sensitive data. Despite its advantages, the distributed nature of FedGNN introduces additional vulnerabilities, particularly backdoor attacks stemming from malicious participants. Although graph backdoor attacks have been explored, the compounded complexity introduced by the combination of GNNs and federated learning has hindered a comprehensive understanding of these attacks, as existing research lacks extensive benchmark coverage and in-depth analysis of critical factors. To address these limitations, we propose Bkd-FedGNN, a benchmark for backdoor attacks on FedGNN. Specifically, Bkd-FedGNN decomposes the graph backdoor attack into trigger generation and injection steps, and extending the attack to the node-level federated setting, resulting in a unified framework that covers both node-level and graph-level classification tasks. Moreover, we thoroughly investigate the impact of multiple critical factors in backdoor attacks on FedGNN. These factors are categorized into global-level and local-level factors, including data distribution, the number of malicious attackers, attack time, overlapping rate, trigger size, trigger type, trigger position, and poisoning rate. Finally, we conduct comprehensive evaluations on 13 benchmark datasets and 13 critical factors, comprising 1,725 experimental configurations for node-level and graph-level tasks from six domains. These experiments encompass over 8,000 individual tests, allowing us to provide a thorough evaluation and insightful observations that advance our understanding of backdoor attacks on FedGNN.The Bkd-FedGNN benchmark is publicly available at https://github.com/usail-hkust/BkdFedGCN.