N-agent ad hoc teamwork (NAHT) is a newly introduced challenge in multi-agent reinforcement learning, where controlled subteams of varying sizes must dynamically collaborate with varying numbers and types of unknown teammates without pre-coordination. The existing learning algorithm (POAM) considers only independent learning for its flexibility in dealing with a changing number of agents. However, independent learning fails to fully capture the inter-agent dynamics essential for effective collaboration. Based on our observation that transformers deal effectively with sequences with varying lengths and have been shown to be highly effective for a variety of machine learning problems, this work introduces a centralized, transformer-based method for N-agent ad hoc teamwork. Our proposed approach incorporates historical observations and actions of all controlled agents, enabling optimal responses to diverse and unseen teammates in partially observable environments. Empirical evaluation on a StarCraft II task demonstrates that MAT-NAHT outperforms POAM, achieving superior sample efficiency and generalization, without auxiliary agent-modeling objectives. Keywords: multi-agent reinforcement learning, ad hoc teamwork, transformers, agent modeling Acknowledgements This work has taken place in the Learning Agents Research Group (LARG) at UT Austin.

Graphical User Interface (GUI) agents show promising capabilities for automating computer-use tasks and facilitating accessibility, but existing interactive benchmarks are mostly English-only, covering web-use or Windows, Linux, and Android environments, but not macOS. macOS is a major OS with distinctive GUI patterns and exclusive applications. To bridge the gaps, we present macOSWorld, the first comprehensive benchmark for evaluating GUI agents on macOS. macOSWorld features 202 multilingual interactive tasks across 30 applications (28 macOS-exclusive), with task instructions and OS interfaces offered in 5 languages (English, Chinese, Arabic, Japanese, and Russian). As GUI agents are shown to be vulnerable to deception attacks, macOSWorld also includes a dedicated safety benchmarking subset. Our evaluation on six GUI agents reveals a dramatic gap: proprietary computer-use agents lead at above 30% success rate, while open-source lightweight research models lag at below 5\%, highlighting the need for macOS domain adaptation. Multilingual benchmarks also expose common weaknesses, especially in Arabic, with a 28.8% average degradation compared to English. Results from safety benchmarking also highlight that deception attacks are more general and demand immediate attention. Project page: https://macos-world.github.io.

Large Language Models (LLMs) have demonstrated strong generative capabilities but remain prone to inconsistencies and hallucinations. We introduce Peer Elicitation Games (PEG), a training-free, game-theoretic framework for aligning LLMs through a peer elicitation mechanism involving a generator and multiple discriminators instantiated from distinct base models. Discriminators interact in a peer evaluation setting, where utilities are computed using a determinant-based mutual information score that provably incentivizes truthful reporting without requiring ground-truth labels. We establish theoretical guarantees showing that each agent, via online learning, achieves sublinear regret in the sense their cumulative performance approaches that of the best fixed truthful strategy in hindsight. Moreover, we prove last-iterate convergence to a truthful Nash equilibrium, ensuring that the actual policies used by agents converge to stable and truthful behavior over time. Empirical evaluations across multiple benchmarks demonstrate significant improvements in factual accuracy. These results position PEG as a practical approach for eliciting truthful behavior from LLMs without supervision or fine-tuning.

How to allocate limited resources to projects that will yield the greatest long-term benefits is a problem that often arises in decision-making under uncertainty. For example, organizations may need to evaluate and select innovation projects with risky returns. Similarly, when allocating resources to research projects, funding agencies are tasked with identifying the most promising proposals based on idiosyncratic criteria. Finally, in participatory budgeting, a local community may need to select a subset of public projects to fund. Regardless of context, agents must estimate the uncertain values of a potentially large number of projects. Developing parsimonious methods to compare these projects, and aggregating agent evaluations so that the overall benefit is maximized, are critical in assembling the best project portfolio. Unlike in standard sorting algorithms, evaluating projects on the basis of uncertain long-term benefits introduces additional complexities. We propose comparison rules based on Quicksort and the Bradley--Terry model, which connects rankings to pairwise "win" probabilities. In our model, each agent determines win probabilities of a pair of projects based on his or her specific evaluation of the projects' long-term benefit. The win probabilities are then appropriately aggregated and used to rank projects. Several of the methods we propose perform better than the two most effective aggregation methods currently available. Additionally, our methods can be combined with sampling techniques to significantly reduce the number of pairwise comparisons. We also discuss how the Bradley--Terry portfolio selection approach can be implemented in practice.

Swarm trajectory optimization problems are a well-recognized class of multi-agent optimal control problems with strong nonlinearity. However, the heuristic nature of needing to set the final time for agents beforehand and the time-consuming limitation of the significant number of iterations prohibit the application of existing methods to large-scale swarm of Unmanned Aerial Vehicles (UAVs) in practice. In this paper, we propose a spatial-temporal trajectory optimization framework that accomplishes multi-UAV consensus based on the Alternating Direction Multiplier Method (ADMM) and uses Differential Dynamic Programming (DDP) for fast local planning of individual UAVs. The introduced framework is a two-level architecture that employs Parameterized DDP (PDDP) as the trajectory optimizer for each UAV, and ADMM to satisfy the local constraints and accomplish the spatial-temporal parameter consensus among all UAVs. This results in a fully distributed algorithm called Distributed Parameterized DDP (D-PDDP). In addition, an adaptive tuning criterion based on the spectral gradient method for the penalty parameter is proposed to reduce the number of algorithmic iterations. Several simulation examples are presented to verify the effectiveness of the proposed algorithm.

The per-label distribution is typically long-tailed (Tarekegn et al., 2021; De Alvis and Seneviratne, 2024): head labels dominate while tail labels appear sporadically. This imbalance is exacerbated in MLC because (i) co-occurring labels make resampling risky, and (ii) metrics like mAP favor head labels. As a result, standard optimizers (Ridnik et al., 2021) often learn head-biased boundaries, achieving high scores while failing on tail labels-problematic for safety-critical applications. In practice the per-label sample counts follow a heavy-tailed distribution: a handful of head labels dominate the data, whereas the vast majority of tail labels appear only sporadically, as shown in Figure 1. This long-tail imbalance (Tarekegn et al., 2021; De Alvis and Seneviratne, 2024) is particularly severe in the multi-label regime because (i) multiple labels co-occur within a single instance, so naïve resampling can destroy cross-label correlations, and (ii) evaluation metrics such as mAP or micro-F1 are disproportionately influenced by head labels, starving tail classes of gradient signal. Consequently, conventional optimizers (Ridnik et al., 2021) that target average loss or accuracy often learn a head-biased decision boundary, yielding high headline scores while silently failing on the tail-an outcome that is unacceptable in safety-critical or comprehensive retrieval scenarios(Barandas et al., 2024).

Autonomous agents rely on automated planning algorithms to achieve their objectives. Simulation-based planning offers a significant advantage over declarative models in modelling complex environments. However, relying solely on a planner that produces a single plan may not be practical, as the generated plans may not always satisfy the agent's preferences. To address this limitation, we introduce $\texttt{FBI}_\texttt{LTL}$, a diverse planner explicitly designed for simulation-based planning problems. $\texttt{FBI}_\texttt{LTL}$ utilises Linear Temporal Logic (LTL) to define semantic diversity criteria, enabling agents to specify what constitutes meaningfully different plans. By integrating these LTL-based diversity models directly into the search process, $\texttt{FBI}_\texttt{LTL}$ ensures the generation of semantically diverse plans, addressing a critical limitation of existing diverse planning approaches that may produce syntactically different but semantically identical solutions. Extensive evaluations on various benchmarks consistently demonstrate that $\texttt{FBI}_\texttt{LTL}$ generates more diverse plans compared to a baseline approach. This work establishes the feasibility of semantically-guided diverse planning in simulation-based environments, paving the way for innovative approaches in realistic, non-symbolic domains where traditional model-based approaches fail.

Recently, a very simple new bilateral negotiation strategy called MiCRO was introduced that does not make use of any kind of opponent modeling or machine learning techniques and that does not require fine-tuning of any parameters. Despite its simplicity, it was shown that MiCRO performs similar to -- or even better than -- most state-of-the-art negotiation strategies. This lead its authors to argue that the benchmark domains on which negotiation algorithms are typically tested may be too simplistic. However, one question that was left open, was how MiCRO could be generalized to multilateral negotiations. In this paper we fill this gap by introducing a multilateral variant of MiCRO. We compare it with the winners of the Automated Negotiating Agents Competitions (ANAC) of 2015, 2017 and 2018 and show that it outperforms them. Furthermore, we perform an empirical game-theoretical analysis to show that our new version of MiCRO forms an empirical Nash equilibrium.

Finding near-optimal solutions for dense multi-agent pathfinding (MAPF) problems in real-time remains challenging even for state-of-the-art planners. To this end, we develop a hybrid framework that integrates a learned heuristic derived from MAGAT, a neural MAPF policy with a graph attention scheme, into a leading search-based algorithm, LaCAM. While prior work has explored learning-guided search in MAPF, such methods have historically underperformed. In contrast, our approach, termed LaGAT, outperforms both purely search-based and purely learning-based methods in dense scenarios. This is achieved through an enhanced MAGAT architecture, a pre-train-then-fine-tune strategy on maps of interest, and a deadlock detection scheme to account for imperfect neural guidance. Our results demonstrate that, when carefully designed, hybrid search offers a powerful solution for tightly coupled, challenging multi-agent coordination problems.

Control-flow hijacking attacks manipulate orchestration mechanisms in multi-agent systems into performing unsafe actions that compromise the system and exfiltrate sensitive information. Recently proposed defenses, such as LlamaFirewall, rely on alignment checks of inter-agent communications to ensure that all agent invocations are "related to" and "likely to further" the original objective. We start by demonstrating control-flow hijacking attacks that evade these defenses even if alignment checks are performed by advanced LLMs. We argue that the safety and functionality objectives of multi-agent systems fundamentally conflict with each other. This conflict is exacerbated by the brittle definitions of "alignment" and the checkers' incomplete visibility into the execution context. LLM-based "agents" equipped with tools for querying APIs, searching the Web, and executing code promise to automate many digital tasks. Popular frameworks like AutoGen (Microsoft, 2025), OpenManus (OpenManus, 2025), CrewAI (CrewAI, 2025), and MetaGPT (MetaGPT, 2025) enable design and deployment of multi-agent systems (MAS). The key principle in MAS is delegation. Given a relatively complex task (e.g., "organize an offsite given team members' calendars, managers' private messages, and Web data about attractions and weather"), MAS can plan how to solve it, delegate sub-tasks to specialized agents, evaluate their responses, and adaptively re-plan if necessary. Delegation splits fulfilling a task into chunks that are (a) hidden within individual agents (e.g., how to access a website or read a file), and (b) joined into the overall plan by an orchestrator who does not observe the execution of sub-tasks, only their results as reported by other agents. Critically, there is no single vantage point in the system where the entire context is visible. This exposes them to indirect prompt injection, or IPI (Greshake et al., 2023), i.e., malicious instructions in the content they ingest (Constantin, 2025; Karliner, 2025; Ravia, 2025; Abu, 2025). Aligning individual agents to resist IPI is not enough. Triedman et al. (2025) demonstrated control-flow hijacking (CFH) attacks that exploit confused-deputy vulnerabilities (Hardy, 1988) in otherwise aligned agents. CFH attacks masquerade as legitimate errors (e.g., failure to parse a file), along with seemingly helpful instructions on how to fix the issue and continue with the user's task. MAS orchestrators receive these instructions from a trusted agent to which they delegated an essential sub-task and rely on them to re-plan the execution and invoke unsafe agents as (indirectly) requested by the attacker.