With the rapidly increasing capabilities and adoption of code agents for AI-assisted coding, safety concerns, such as generating or executing risky code, have become significant barriers to the real-world deployment of these agents. To provide comprehensive and practical evaluations on the safety of code agents, we propose RedCode, a benchmark for risky code execution and generation: (1) RedCode-Exec provides challenging prompts that could lead to risky code execution, aiming to evaluate code agents' ability to recognize and handle unsafe code. We provide a total of 4,050 risky test cases in Python and Bash tasks with diverse input formats including code snippets and natural text. They covers 25 types of critical vulnerabilities spanning 8 domains (e.g., websites, file systems). We provide Docker environments and design corresponding evaluation metrics to assess their execution results. (2) RedCode-Gen provides 160 prompts with function signatures and docstrings as input to assess whether code agents will follow instructions to generate harmful code or software. Our empirical findings, derived from evaluating three agent frameworks based on 19 LLMs, provide insights into code agents' vulnerabilities. For instance, evaluations on RedCode-Exec show that agents are more likely to reject executing risky operations on the operating system, but are less likely to reject executing technically buggy code, indicating high risks. Risky operations described in natural text lead to a lower rejection rate than those in code format. Additionally, evaluations on RedCode-Gen show that more capable base models and agents with stronger overall coding abilities, such as GPT4, tend to produce more sophisticated and effective harmful software. Our findings highlight the need for stringent safety evaluations for diverse code agents. Our dataset and code are available at https://github.com/AI-secure/RedCode.

We study the problem of multi-agent multi-armed bandits with adversarial corruption in a heterogeneous setting, where each agent accesses a subset of arms. The adversary can corrupt the reward observations for all agents. Agents share these corrupted rewards with each other, and the objective is to maximize the cumulative total reward of all agents (and not be misled by the adversary). We propose a multi-agent cooperative learning algorithm that is robust to adversarial corruptions. For this newly devised algorithm, we demonstrate that an adversary with an unknown corruption budget $C$ only incurs an additive $O((L / L_{\min}) C)$ term to the standard regret of the model in non-corruption settings, where $L$ is the total number of agents, and $L_{\min}$ is the minimum number of agents with mutual access to an arm. As a side-product, our algorithm also improves the state-of-the-art regret bounds when reducing to both the single-agent and homogeneous multi-agent scenarios, tightening multiplicative $K$ (the number of arms) and $L$ (the number of agents) factors, respectively.

Mean field games (MFGs) tractably model behavior in large agent populations. The literature on learning MFG equilibria typically focuses on finding Nash equilibria (NE), which assume perfectly rational agents and are hence implausible in many realistic situations. To overcome these limitations, we incorporate bounded rationality into MFGs by leveraging the well-known concept of quantal response equilibria (QRE). Two novel types of MFG QRE enable the modeling of large agent populations where individuals only noisily estimate the true objective. We also introduce a second source of bounded rationality to MFGs by restricting agents' planning horizon. The resulting novel receding horizon (RH) MFGs are combined with QRE and existing approaches to model different aspects of bounded rationality in MFGs. We formally define MFG QRE and RH MFGs and compare them to existing equilibrium concepts such as entropy-regularized NE. Subsequently, we design generalized fixed point iteration and fictitious play algorithms to learn QRE and RH equilibria. After a theoretical analysis, we give different examples to evaluate the capabilities of our learning algorithms and outline practical differences between the equilibrium concepts.

This study investigates the use of generative AI and multi-agent systems to provide automatic feedback in educational contexts, particularly for student constructed responses in science assessments. The research addresses a key gap in the field by exploring how multi-agent systems, called AutoFeedback, can improve the quality of GenAI-generated feedback, overcoming known issues such as over-praise and over-inference that are common in single-agent large language models (LLMs). The study developed a multi-agent system consisting of two AI agents: one for generating feedback and another for validating and refining it. The system was tested on a dataset of 240 student responses, and its performance was compared to that of a single-agent LLM. Results showed that AutoFeedback significantly reduced the occurrence of over-praise and over-inference errors, providing more accurate and pedagogically sound feedback. The findings suggest that multi-agent systems can offer a more reliable solution for generating automated feedback in educational settings, highlighting their potential for scalable and personalized learning support. These results have important implications for educators and researchers seeking to leverage AI in formative assessments, offering a pathway to more effective feedback mechanisms that enhance student learning outcomes.

In an era where artificial intelligence (AI) is reshaping countless fields, the research community of social sciences needs to adapt to the changes posed by these technologies [1, 2]. In particular, data quality and authenticity play a significant role in social sciences [3], where the conclusions drawn rely heavily on data collected, for instance, from surveys. There are many traditional ways of gathering data, such as public datasets or private surveys, but AI has led to innovative approaches, like using agent-based models (ABMs). In recent years, the use of this paradigm has gained significant attention across a variety of fields, from economics and social sciences to artificial intelligence and computational biology [4, 5, 6]. ABMs allow researchers to simulate complex situations by modeling the behaviors and interactions of individual agents within a given environment [7]. These models provide a powerful way to understand emergent phenomena--such as market dynamics, social behaviors, or ecological systems--that arise from the independent actions and interactions of individual agents, each following its own set of rules. In spite of their flexibility, these models face some limitations, particularly when dealing with complex environments. One of the main challenges is that the agents' behaviors are programmed by the modeler based on assumptions or simplified rules. This rigid structure limits the ability to account for the full range of possible interactions that can emerge in real-world scenarios.

The systems of large number (>10) of agents, hereafter referred to as a multi-agent system, are crucial in a wide range of autonomy applications, including swarm robotics [1] and fleets of autonomous vehicles [2]. Inspired by collective behaviors observed in nature such as fish schools and bird flocks, these systems aim to achieve collective goals through the interaction among individual agents using a set of decentralized rules. Analytical flocking models such as Reynolds model [3] or Vicsek model [4] replicate collective behaviors observed in nature, but these models require precise localization which is rarely possible in the real-world applications. Therefore, real-time prediction of collective behavior, like how and when agents will achieve a collective goal, is essential for adapting the local rules and controlling multi-agent systems in a real-world environment [5, 6] as illustrated in Figure 1. This prediction is valuable in competitive settings like swarm herding [7], where understanding the system dynamics of adversarial agents can enhance strategic control.

Understanding how individual agents make strategic decisions within collectives is important for advancing fields as diverse as economics, neuroscience, and multi-agent systems. Two complementary approaches can be integrated to this end. The Active Inference framework (AIF) describes how agents employ a generative model to adapt their beliefs about and behaviour within their environment. Game theory formalises strategic interactions between agents with potentially competing objectives. To bridge the gap between the two, we propose a factorisation of the generative model whereby each agent maintains explicit, individual-level beliefs about the internal states of other agents, and uses them for strategic planning in a joint context. We apply our model to iterated general-sum games with 2 and 3 players, and study the ensemble effects of game transitions, where the agents' preferences (game payoffs) change over time. This non-stationarity, beyond that caused by reciprocal adaptation, reflects a more naturalistic environment in which agents need to adapt to changing social contexts. Finally, we present a dynamical analysis of key AIF quantities: the variational free energy (VFE) and the expected free energy (EFE) from numerical simulation data. The ensemble-level EFE allows us to characterise the basins of attraction of games with multiple Nash Equilibria under different conditions, and we find that it is not necessarily minimised at the aggregate level. By integrating AIF and game theory, we can gain deeper insights into how intelligent collectives emerge, learn, and optimise their actions in dynamic environments, both cooperative and non-cooperative.

This study investigates the efficacy of Multi-Agent Systems in eliciting cross-agent communication and enhancing collective intelligence through group decision-making in a decentralized setting. Unlike centralized mechanisms, where a fixed hierarchy governs social choice, decentralized group decision-making allows agents to engage in joint deliberation. Our research focuses on the dynamics of communication and decision-making within various social choice methods. By applying different voting rules in various environments, we find that moderate decision flexibility yields better outcomes. Additionally, exploring the linguistic features of agent-to-agent conversations reveals indicators of effective collaboration, offering insights into communication patterns that facilitate or hinder collaboration. Finally, we propose various methods for determining the optimal stopping point in multi-agent collaborations based on linguistic cues. Our findings contribute to a deeper understanding of how decentralized decision-making and group conversation shape multi-agent collaboration, with implications for the design of more effective MAS environments.

Building a distributed spatial awareness within a swarm of locally sensing and communicating robots enables new swarm algorithms. We use local observations by robots of each other and Gaussian Belief Propagation message passing combined with continuous swarm movement to build a global and distributed swarm-centric frame of reference. With low bandwidth and computation requirements, this shared reference frame allows new swarm algorithms. We characterise the system in simulation and demonstrate two example algorithms.

As modern web services increasingly rely on REST APIs, their thorough testing has become crucial. Furthermore, the advent of REST API specifications such as the OpenAPI Specification has led to the emergence of many black-box REST API testing tools. However, these tools often focus on individual test elements in isolation (e.g., APIs, parameters, values), resulting in lower coverage and less effectiveness in detecting faults (i.e., 500 response codes). To address these limitations, we present AutoRestTest, the first black-box framework to adopt a dependency-embedded multi-agent approach for REST API testing, integrating Multi-Agent Reinforcement Learning (MARL) with a Semantic Property Dependency Graph (SPDG) and Large Language Models (LLMs). Our approach treats REST API testing as a separable problem, where four agents -- API, dependency, parameter, and value -- collaborate to optimize API exploration. LLMs handle domain-specific value restrictions, the SPDG model simplifies the search space for dependencies using a similarity score between API operations, and MARL dynamically optimizes the agents' behavior. Evaluated on 12 real-world REST services, AutoRestTest outperforms the four leading black-box REST API testing tools, including those assisted by RESTGPT (which augments realistic test inputs using LLMs), in terms of code coverage, operation coverage, and fault detection. Notably, AutoRestTest is the only tool able to identify an internal server error in Spotify. Our ablation study underscores the significant contributions of the agent learning, SPDG, and LLM components.