Vision-and-Language navigation on websites requires agents to navigate target webpages and answer questions based on human instructions. Current web agents primarily leverage Large Language Models (LLMs) for semantic understanding and reasoning, but still suffer from limited navigation performance and slow inference speed. Constructing a global map across webpages can effectively enhance both navigation accuracy and efficiency, however, this is challenged by the open structure of web navigation graphs and the dynamic nature of web layouts. In this paper, we propose ATLAS: Adaptive Topological Layout And Semantic mapping, a framework that adaptively constructs a time-varying, unbounded topological map across webpages and unifies heterogeneous elements through semantic representation. This enables both global path planning and local element selection for web-based navigation and question answering. As a lightweight approach, ATLAS significantly outperforms existing state-of-the-art methods on the WebVLN benchmark with a 10% improvement in success rate, and achieves the highest average task success rate on both the Mind2Web and WebArena benchmarks.

Recent breakthroughs in large language model-driven autonomous agents have revealed that multi-agent collaboration often surpasses each individual through collective reasoning. Inspired by the neural scaling law—increasing neurons enhances performance, this study explores whether the continuous addition of collaborative agents can yield similar benefits. Technically, we utilize directed acyclic graphs to organize agents into a multi-agent collaboration network (MacNet), upon which their interactive reasoning is topologically orchestrated for autonomous task solving. Extensive evaluations reveal that it effectively supports collaboration among over a thousand agents, with irregular topologies outperforming regular ones. We also identify a collaborative scaling law—the overall performance follows a logistic growth pattern as agents scale, with collaborative emergence occurring earlier than traditional neural emergence. We speculate this may be because scaling agents catalyzes their multidimensional considerations during interactive reflection and refinement, thereby producing more comprehensive artifacts. The code is available at https://github.com/OpenBMB/ChatDev/tree/macnet.

Vision-and-Language Navigation (VLN) is a critical task that enables robots to comprehend human instructions. The premise of current VLN tasks is built on the human’s familiarity with the environment structure, guiding the agent to complete the navigation task with explicit instructions, referred to as Human Guidance VLN (HG-VLN). However, real-world scenarios often involve humans unfamiliar with new environments, relying on agents to assist with navigation. In such tasks, humans can only provide destination-related information, requiring the agent to perform the path planning. We term this scenario Human Assistant VLN (HA-VLN). HA-VLN poses greater demands on the agent, therefore, we have restructured the classic Room to Room (R2R) dataset to introduce the Room to Room Assistant (R2RA) dataset, tailored for HA-VLN tasks. To address the challenges existing methods face when processing HA-VLN task instructions, we propose HAVEN: Human Assistant Vision-and- Language Navigation Evolution Network. This network integrates a Large Language Model (LLM) with an embedded memory system, achieving the paradigm shift from mainstream VLN methods to the HA-VLN task without requiring additional information. Our experiments demonstrate that algorithms incorporating HAVEN can achieve higher success rates in reaching destinations, shorter path selection, and lower navigation error rates in HA-VLN tasks. HAVEN can be integrated as an end-to-end module into any VLN method. The code and dataset is available at: https://github.com/longziyu/R2RA-Dataset

Large language models (LLMs) have achieved remarkable results across diverse downstream tasks, but their monolithic nature restricts scalability and efficiency in complex problem-solving. While recent research explores multi-agent collaboration among LLMs, most approaches rely on static organizational structures that struggle to adapt as task complexity and agent numbers grow, resulting in coordination overhead and inefficiencies. To this end, we propose a puppeteer-style paradigm for LLM-based multi-agent collaboration, where a centralized orchestrator ("puppeteer") dynamically directs agents ("puppets") in response to evolving task states. This orchestrator is trained via reinforcement learning to adaptively sequence and prioritize agents, enabling flexible and evolvable collective reasoning. Experiments on closed- and open-domain scenarios show that this method achieves superior performance with reduced computational costs. Analyses further reveal that the key improvements consistently stem from the emergence of more compact, cyclic reasoning structures under the orchestrator’s evolution. Our code is available at https://github.com/OpenBMB/ChatDev/tree/puppeteer.

Chen Qian, Yufan Dang, Jiahao Li, Wei Liu, Zihao Xie, YiFei Wang, Weize Chen, Cheng Yang, Xin Cong, Xiaoyin Che, Zhiyuan Liu, Maosong Sun. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Large Language Model Multi-Agent Systems (LLM-MAS) have greatly progressed in solving complex tasks. It communicates among agents within the system to collaboratively solve tasks, under the premise of shared information. However, when agents' collaborations are leveraged to perform multi-person tasks, a new challenge arises due to information asymmetry, since each agent can only access the information of its human user. Previous MAS struggle to complete tasks under this condition. To address this, we propose a new MAS paradigm termed iAgents, which denotes Informative Multi-Agent Systems. In iAgents, the human social network is mirrored in the agent network, where agents proactively exchange human information necessary for task resolution, thereby overcoming information asymmetry. iAgents employs a novel agent reasoning mechanism, InfoNav, to navigate agents' communication towards effective information exchange. Together with InfoNav, iAgents organizes human information in a mixed memory to provide agents with accurate and comprehensive information for exchange. Additionally, we introduce InformativeBench, the first benchmark tailored for evaluating LLM agents' task-solving ability under information asymmetry. Experimental results show that iAgents can collaborate within a social network of 140 individuals and 588 relationships, autonomously communicate over 30 turns, and retrieve information from nearly 70,000 messages to complete tasks within 3 minutes.

Industrial control networks are frequent targets of cyber attacks, calling for autonomous defense strategies to combat threats effectively and promptly. We propose to leverage Reinforcement Learning (RL) to train defenders how to select the most appropriate actions in response to attackers’ behavior. We model the defender and attacker as agents in RL environments and define action space and reward function. We focus on evaluating value-based and policy gradient-based RL algorithms and propose enhancements to Proximal Policy Optimization (PPO) method through normalization. We then leverage CybORG, a simulation tool to study how the RL algorithms perform to achieve autonomous defense. Our extensive experiments reveal that the proposed normalized PPO outperforms other models in terms of stability, robustness, and average rewards. The results also demonstrate the effectiveness of the defender in protecting the operational server upon changing the attacker’s position. Additionally, increasing the number of attackers had a significant impact on policy-based algorithms, particularly PPO, with decreased reward values highlighting the increased vulnerability of the network.

Facing the global crisis brought by COVID-19, many countries have adopted social distancing or stay-at-home measures to restrict individual mobility to control the virus. Mean-while, the availability of anonymized and aggregated mobility data provides an opportunity to obtain a deeper understanding of the impact of these measures. In this paper, we utilize an open mobility dataset called Community Mobility Report published by Google on the Internet and other external data sources (e.g., statistics on daily confirmed cases, demographics, etc.) to quantitatively characterize people’s collective responses and model it with a proposed metric called Lockdown Stringency Score (LSS) after the lockdown measures have been taken. Then, by investigating the correlations between LSS and the increase of new confirmed cases across different regions and countries in the world, we explore how people’s collective response in terms of mobility pattern changes affects the control of the virus. The analysis results show that lockdown and social distancing measures do have a positive impact on virus control, and the restriction on different types of Point-of-Interests (PoIs) has different weights (significance) in terms of virus control effectiveness. These results reveal important insights and implication on public health policy making, such as the phased start of the lockdown or reopen of the economy.