Modern human labor is characterized by specialization; we train for years and develop particular tools that allow us to perform well across a variety of tasks. In addition, AI agents have been specialized for domains such as software engineering, web navigation, and workflow automation. However, this results in agents that are good for one thing but fail to generalize beyond their intended scope. One reason for this is that agent developers provide a highly specialized set of tools or make architectural decisions optimized for a specific use case or benchmark. In this work, we ask the question: what is the minimal set of general tools that can be used to achieve high performance across a diverse set of tasks? Our answer is OpenHands-Versa, a generalist agent built with a modest number of general tools: code editing and execution, web search, as well as multimodal web browsing and file access. Importantly, OpenHands-Versa demonstrates superior or competitive performance over leading specialized agents across three diverse and challenging benchmarks: SWE-Bench Multimodal, GAIA, and The Agent Company, outperforming the best-performing previously published results with absolute improvements in success rate of 9.1, 1.3, and 9.1 points respectively. Further, we show how existing state-of-the-art multi-agent systems fail to generalize beyond their target domains. These results demonstrate the feasibility of developing a generalist agent to solve diverse tasks and establish OpenHands-Versa as a strong baseline for future research.

Privacy policies inform users about data collection and usage, yet their complexity limits accessibility for diverse populations. Existing Privacy Policy Question Answering (QA) systems exhibit performance disparities across English dialects, disadvantaging speakers of non-standard varieties. We propose a novel multi-agent framework inspired by human-centered design principles to mitigate dialectal biases. Our approach integrates a Dialect Agent, which translates queries into Standard American English (SAE) while preserving dialectal intent, and a Privacy Policy Agent, which refines predictions using domain expertise. Unlike prior approaches, our method does not require retraining or dialect-specific fine-tuning, making it broadly applicable across models and domains. Evaluated on PrivacyQA and PolicyQA, our framework improves GPT-4o-mini's zero-shot accuracy from 0.394 to 0.601 on PrivacyQA and from 0.352 to 0.464 on PolicyQA, surpassing or matching few-shot baselines without additional training data. These results highlight the effectiveness of structured agent collaboration in mitigating dialect biases and underscore the importance of designing NLP systems that account for linguistic diversity to ensure equitable access to privacy information.

Counterfactual inference is a powerful tool for analysing and evaluating autonomous agents, but its application to language model (LM) agents remains challenging. Existing work on counterfactuals in LMs has primarily focused on token-level counterfactuals, which are often inadequate for LM agents due to their open-ended action spaces. Unlike traditional agents with fixed, clearly defined action spaces, the actions of LM agents are often implicit in the strings they output, making their action spaces difficult to define and interpret. Furthermore, the meanings of individual tokens can shift depending on the context, adding complexity to token-level reasoning and sometimes leading to biased or meaningless counterfactuals. We introduce \emph{Abstract Counterfactuals}, a framework that emphasises high-level characteristics of actions and interactions within an environment, enabling counterfactual reasoning tailored to user-relevant features. Our experiments demonstrate that the approach produces consistent and meaningful counterfactuals while minimising the undesired side effects of token-level methods. We conduct experiments on text-based games and counterfactual text generation, while considering both token-level and latent-space interventions.

Evaluating the security of multi-agent systems (MASs) powered by large language models (LLMs) is challenging, primarily because of the systems' complex internal dynamics and the evolving nature of LLM vulnerabilities. Traditional attack graph (AG) methods often lack the specific capabilities to model attacks on LLMs. This paper introduces AI-agent application Threat assessment with Attack Graphs (ATAG), a novel framework designed to systematically analyze the security risks associated with AI-agent applications. ATAG extends the MulVAL logic-based AG generation tool with custom facts and interaction rules to accurately represent AI-agent topologies, vulnerabilities, and attack scenarios. As part of this research, we also created the LLM vulnerability database (LVD) to initiate the process of standardizing LLM vulnerabilities documentation. To demonstrate ATAG's efficacy, we applied it to two multi-agent applications. Our case studies demonstrated the framework's ability to model and generate AGs for sophisticated, multi-step attack scenarios exploiting vulnerabilities such as prompt injection, excessive agency, sensitive information disclosure, and insecure output handling across interconnected agents. ATAG is an important step toward a robust methodology and toolset to help understand, visualize, and prioritize complex attack paths in multi-agent AI systems (MAASs). It facilitates proactive identification and mitigation of AI-agent threats in multi-agent applications.

Web agents for online shopping have shown great promise in automating user interactions across e-commerce platforms. Benchmarks for assessing such agents do not reflect the complexity of real-world shopping scenarios, as they often consist of overly simple queries with deterministic paths, such as "Find iPhone 15." Real shopping scenarios are inherently more layered, involving multi-dimensional product attributes, search filters, and user-specific sorting preferences. To address this gap, we introduce DeepShop, a benchmark designed to evaluate web agents in complex and realistic online shopping environments. DeepShop comprises three key components. (1) Query diversity evolution: Starting from real user queries, we generate diverse queries across five popular online shopping domains. (2) Query complexity evolution: We further evolve these queries to increase complexity, considering product attributes, search filters, and sorting preferences, and classify them into three levels: easy, medium, and hard, based on the number of evolutions. (3) Fine-grained and holistic evaluation: We propose an automated evaluation framework that assesses agent performance in terms of fine-grained aspects (product attributes, search filters, and sorting preferences) and reports the overall success rate through holistic evaluation. We conduct a systematic evaluation of retrieval-augmented generation (RAG) methods, web agents, and deep research systems. Results show that RAG struggles with complex queries due to its lack of web interaction, while other methods face significant challenges with filters and sorting preferences, leading to low overall success rates. We also perform cross-category, complexity-based evaluations and error analyses to support the advancement of deep research shopping agents.

Large Language Models (LLMs) have become foundational to modern AI agent systems, enabling autonomous agents to reason and plan. In most existing systems, inter-agent communication relies primarily on natural language. While this design supports interpretability and human oversight, we argue that it introduces fundamental limitations in agent-to-agent coordination. The semantic space of natural language is structurally misaligned with the high-dimensional vector spaces in which LLMs operate, resulting in information loss and behavioral drift. Beyond surface-level inefficiencies, we highlight a deeper architectural limitation: current LLMs were not trained with the objective of supporting agentic behavior. As such, they lack mechanisms for modeling role continuity, task boundaries, and multi-agent dependencies. The standard next-token prediction paradigm fails to support the structural alignment required for robust, scalable agent coordination. Based on this, we argue that two core questions deserve careful examination: first, given that AI agents fundamentally operate in high-dimensional vector spaces, should they rely on a language system originally designed for human cognition as their communication medium? Second, should we consider developing a new model construction paradigm that builds models from the ground up to natively support structured communication, shared intentionality, and task alignment in multi-role, multi-agent environments? This paper calls for a reconsideration not only of how agents should communicate, but also of what it fundamentally means to train a model that natively supports multi-agent coordination and communication.

Large Language Models (LLMs) have achieved remarkable success across diverse natural language processing tasks, yet their deployment in real-world applications is hindered by fixed knowledge cutoffs and difficulties in generating controllable, accurate outputs in a single inference. Multi-agent systems (MAS) built from specialized LLM agents offer a promising solution, enabling dynamic collaboration and iterative reasoning. However, optimizing these systems remains a challenge, as conventional methods such as prompt engineering and supervised fine-tuning entail high engineering overhead and limited adaptability. Reinforcement learning (RL), particularly multi-agent reinforcement learning (MARL), provides a scalable framework by refining agent policies based on system-level feedback. Nevertheless, existing MARL algorithms--such as Multi-Agent Proximal Policy Optimization (MAPPO)--rely on Critic networks, which can cause training instability and increase computational burden. To address these limitations and target the prototypical Multi-Agent Search System (MASS), we propose Multi-Agent Heterogeneous Group Policy Optimization (MHGPO), a novel Critic-free algorithm that guides policy updates by estimating relative reward advantages across heterogeneous groups of rollouts. MHGPO eliminates the need for Critic networks, enhancing stability and reducing computational overhead. Additionally, we introduce three group rollout sampling strategies that trade off between efficiency and effectiveness. Experiments on a multi-agent LLM-based search system demonstrate that MHGPO consistently outperforms MAPPO in both task performance and computational efficiency, without requiring warm-up, underscoring its potential for stable and scalable optimization of complex LLM-based MAS.

Abstract--Self-organizing networks face challenges from complex parameter interdependencies and conflicting objec - tives. This study introduces two compositional learning ap - proaches--Compositional Deep Reinforcement Learning (CDR L) and Compositional Predictive Decision-Making (CPDM)--and evaluates their performance under training time and safety constraints in multi-agent systems. We propose a modular, t wo-tier framework with cell-level and cell-pair-level agents to manage heterogeneous agent granularities while reducing model co mplex-ity. Numerical simulations reveal a 37.2% reduction in handover failures, along with improved throughput and latency, outp er-forming conventional multi-agent deep reinforcement lear ning approaches. The approach also demonstrates superior scala bility, faster convergence, higher sample efficiency, and safer tra ining in large-scale self-organizing networks. Self-organizing networks (SON) is a key enabler of autonomous networks, leveraging mechanisms like mobility ro - bustness optimization (MRO) and mobility load balancing (MLB) [1] to dynamically optimize network control parameters using key performance indicators (KPIs).

We consider the problem of indoor building-scale social navigation, where the robot must reach a point goal as quickly as possible without colliding with humans who are freely moving around. Factors such as varying crowd densities, unpredictable human behavior, and the constraints of indoor spaces add significant complexity to the navigation task, necessitating a more advanced approach. We propose a modular navigation framework that leverages the strengths of both classical methods and deep reinforcement learning (DRL). Our approach employs a global planner to generate waypoints, assigning soft costs around anticipated pedestrian locations, encouraging caution around potential future positions of humans. Simultaneously, the local planner, powered by DRL, follows these waypoints while avoiding collisions. The combination of these planners enables the agent to perform complex maneuvers and effectively navigate crowded and constrained environments while improving reliability. Many existing studies on social navigation are conducted in simplistic or open environments, limiting the ability of trained models to perform well in complex, real-world settings. To advance research in this area, we introduce a new 2D benchmark designed to facilitate development and testing of social navigation strategies in indoor environments. We benchmark our method against traditional and RL-based navigation strategies, demonstrating that our approach outperforms both.

Computer-Use Agents (CUAs) with full system access enable powerful task automation but pose significant security and privacy risks due to their ability to manipulate files, access user data, and execute arbitrary commands. While prior work has focused on browser-based agents and HTML-level attacks, the vulnerabilities of CUAs remain underexplored. In this paper, we investigate Visual Prompt Injection (VPI) attacks, where malicious instructions are visually embedded within rendered user interfaces, and examine their impact on both CUAs and Browser-Use Agents (BUAs). We propose VPI-Bench, a benchmark of 306 test cases across five widely used platforms, to evaluate agent robustness under VPI threats. Each test case is a variant of a web platform, designed to be interactive, deployed in a realistic environment, and containing a visually embedded malicious prompt. Our empirical study shows that current CUAs and BUAs can be deceived at rates of up to 51% and 100%, respectively, on certain platforms. The experimental results also indicate that system prompt defenses offer only limited improvements. These findings highlight the need for robust, context-aware defenses to ensure the safe deployment of multimodal AI agents in real-world environments. The code and dataset are available at: https://github.com/cua-framework/agents