Large language model based multi-agent systems have demonstrated significant potential in social simulation and complex task resolution domains. However, current frameworks face critical challenges in system architecture design, cross-domain generalizability, and performance guarantees, particularly as task complexity and number of agents increases. We introduces AgentGroupChat-V2, a novel framework addressing these challenges through three core innovations: (1) a divide-and-conquer fully parallel architecture that decomposes user queries into hierarchical task forest structures enabling dependency management and distributed concurrent processing. (2) an adaptive collaboration engine that dynamically selects heterogeneous LLM combinations and interaction modes based on task characteristics. (3) agent organization optimization strategies combining divide-and-conquer approaches for efficient problem decomposition. Extensive experiments demonstrate AgentGroupChat-V2's superior performance across diverse domains, achieving 91.50% accuracy on GSM8K (exceeding the best baseline by 5.6 percentage points), 30.4% accuracy on competition-level AIME (nearly doubling other methods), and 79.20% pass@1 on HumanEval. Performance advantages become increasingly pronounced with higher task difficulty, particularly on Level 5 MATH problems where improvements exceed 11 percentage points compared to state-of-the-art baselines. These results confirm that AgentGroupChat-V2 provides a comprehensive solution for building efficient, general-purpose LLM multi-agent systems with significant advantages in complex reasoning scenarios. Code is available at https://github.com/MikeGu721/AgentGroupChat-V2.

This work focuses on the credit assignment problem in cooperative multi-agent reinforcement learning (MARL). Sharing the global advantage among agents often leads to suboptimal policy updates as it fails to account for the distinct contributions of agents. Although numerous methods consider global or individual contributions for credit assignment, a detailed analysis at the coalition level remains lacking in many approaches. This work analyzes the over-updating problem during multi-agent policy updates from a coalition-level perspective. To address this issue, we propose a credit assignment method called Coalitional Rational Advantage Decomposition (CORA). CORA evaluates coalitional advantages via marginal contributions from all possible coalitions and decomposes advantages using the core solution from cooperative game theory, ensuring coalitional rationality. To reduce computational overhead, CORA employs random coalition sampling. Experiments on matrix games, differential games, and multi-agent collaboration benchmarks demonstrate that CORA outperforms strong baselines, particularly in tasks with multiple local optima. These findings highlight the importance of coalition-aware credit assignment for improving MARL performance.

Developing agents capable of exploring, planning and learning in complex open-ended environments is a grand challenge in artificial intelligence (AI). Hierarchical reinforcement learning (HRL) offers a promising solution to this challenge by discovering and exploiting the temporal structure within a stream of experience. The strong appeal of the HRL framework has led to a rich and diverse body of literature attempting to discover a useful structure. However, it is still not clear how one might define what constitutes good structure in the first place, or the kind of problems in which identifying it may be helpful. This work aims to identify the benefits of HRL from the perspective of the fundamental challenges in decision-making, as well as highlight its impact on the performance trade-offs of AI agents. Through these benefits, we then cover the families of methods that discover temporal structure in HRL, ranging from learning directly from online experience to offline datasets, to leveraging large language models (LLMs). Finally, we highlight the challenges of temporal structure discovery and the domains that are particularly well-suited for such endeavours.

Reinforcement Learning (RL) agents often exhibit learning behaviors that are not intuitively interpretable by human observers, which can result in suboptimal feedback in collaborative teaching settings. Yet, how humans perceive and interpret RL agent's learning behavior is largely unknown. In a bottom-up approach with two experiments, this work provides a data-driven understanding of the factors of human observers' understanding of the agent's learning process. A novel, observation-based paradigm to directly assess human inferences about agent learning was developed. In an exploratory interview study (\textit{N}=9), we identify four core themes in human interpretations: Agent Goals, Knowledge, Decision Making, and Learning Mechanisms. A second confirmatory study (\textit{N}=34) applied an expanded version of the paradigm across two tasks (navigation/manipulation) and two RL algorithms (tabular/function approximation). Analyses of 816 responses confirmed the reliability of the paradigm and refined the thematic framework, revealing how these themes evolve over time and interrelate. Our findings provide a human-centered understanding of how people make sense of agent learning, offering actionable insights for designing interpretable RL systems and improving transparency in Human-Robot Interaction.

Are world models a necessary ingredient for flexible, goal-directed behaviour, or is model-free learning sufficient? We provide a formal answer to this question, showing that any agent capable of generalizing to multi-step goal-directed tasks must have learned a predictive model of its environment. We show that this model can be extracted from the agent's policy, and that increasing the agents performance or the complexity of the goals it can achieve requires learning increasingly accurate world models. This has a number of consequences: from developing safe and general agents, to bounding agent capabilities in complex environments, and providing new algorithms for eliciting world models from agents.

Virtual environments are essential to AI agent research. Existing environments for LLM agent research typically focus on either physical task solving or social simulation, with the former oversimplifying agent individuality and social dynamics, and the latter lacking physical grounding of social behaviors. We introduce IndoorWorld, a heterogeneous multi-agent environment that tightly integrates physical and social dynamics. By introducing novel challenges for LLM-driven agents in orchestrating social dynamics to influence physical environments and anchoring social interactions within world states, IndoorWorld opens up possibilities of LLM-based building occupant simulation for architectural design. We demonstrate the potential with a series of experiments within an office setting to examine the impact of multi-agent collaboration, resource competition, and spatial layout on agent behavior.

Understanding how complex societal behaviors emerge from individual cognition and interactions requires both high-fidelity modeling of human behavior and large-scale simulations. Traditional agent-based models (ABMs) have been employed to study these dynamics for decades, but are constrained by simplified agent behaviors that fail to capture human complexity. Recent advances in large language models (LLMs) offer new opportunities by enabling agents to exhibit sophisticated social behaviors that go beyond rule-based logic, yet face significant scaling challenges. Here we present Light Society, an agent-based simulation framework that advances both fronts, efficiently modeling human-like societies at planetary scale powered by LLMs. Light Society formalizes social processes as structured transitions of agent and environment states, governed by a set of LLM-powered simulation operations, and executed through an event queue. This modular design supports both independent and joint component optimization, supporting efficient simulation of societies with over one billion agents. Large-scale simulations of trust games and opinion propagation--spanning up to one billion agents--demonstrate Light Society's high fidelity and efficiency in modeling social trust and information diffusion, while revealing scaling laws whereby larger simulations yield more stable and realistic emergent behaviors.

Learning to cooperate in distributed partially observable environments with no communication abilities poses significant challenges for multi-agent deep reinforcement learning (MARL). This paper addresses key concerns in this domain, focusing on inferring state representations from individual agent observations and leveraging these representations to enhance agents' exploration and collaborative task execution policies. To this end, we propose a novel state modelling framework for cooperative MARL, where agents infer meaningful belief representations of the non-observable state, with respect to optimizing their own policies, while filtering redundant and less informative joint state information. Building upon this framework, we propose the MARL SMPE algorithm. In SMPE, agents enhance their own policy's discriminative abilities under partial observability, explicitly by incorporating their beliefs into the policy network, and implicitly by adopting an adversarial type of exploration policies which encourages agents to discover novel, high-value states while improving the discriminative abilities of others. Experimentally, we show that SMPE outperforms state-of-the-art MARL algorithms in complex fully cooperative tasks from the MPE, LBF, and RWARE benchmarks.

Open multi-agent systems are increasingly important in modeling real-world applications, such as smart grids, swarm robotics, etc. In this paper, we aim to investigate a recently proposed problem for open multi-agent systems, referred to as n-agent ad hoc teamwork (NAHT), where only a number of agents are controlled. Existing methods tend to be based on heuristic design and consequently lack theoretical rigor and ambiguous credit assignment among agents. To address these limitations, we model and solve NAHT through the lens of cooperative game theory. More specifically, we first model an open multi-agent system, characterized by its value, as an instance situated in a space of cooperative games, generated by a set of basis games. We then extend this space, along with the state space, to accommodate dynamic scenarios, thereby characterizing NAHT. Exploiting the justifiable assumption that basis game values correspond to a sequence of n-step returns with different horizons, we represent the state values for NAHT in a form similar to $λ$-returns. Furthermore, we derive Shapley values to allocate state values to the controlled agents, as credits for their contributions to the ad hoc team. Different from the conventional approach to shaping Shapley values in an explicit form, we shape Shapley values by fulfilling the three axioms uniquely describing them, well defined on the extended game space describing NAHT. To estimate Shapley values in dynamic scenarios, we propose a TD($λ$)-like algorithm. The resulting reinforcement learning (RL) algorithm is referred to as Shapley Machine. To our best knowledge, this is the first time that the concepts from cooperative game theory are directly related to RL concepts. In experiments, we demonstrate the effectiveness of Shapley Machine and verify reasonableness of our theory.

Large language model (LLM) have become mainstream methods in the field of sarcasm detection. However, existing LLM methods face challenges in irony detection, including: 1. single-perspective limitations, 2. insufficient comprehensive understanding, and 3. lack of interpretability . This paper introduces the Collaborative Agent Framework for Irony ( CAF-I), an LLM-driven multi-agent system designed to overcome these issues. CAF-I employs specialized agents for Context, Semantics, and Rhetoric, which perform multidimensional analysis and engage in interactive collaborative optimization. A Decision Agent then consolidates these perspectives, with a Refinement Evaluator Agent providing conditional feedback for optimization. Experiments on benchmark datasets establish CAF-I's state-of-the-art zero-shot performance. Achieving SOTA on the vast majority of metrics, CAF-I reaches an average Macro-F1 of 76.31%, a 4.98% absolute improvement over the strongest prior baseline. This success is attained by its effective simulation of human-like multi-perspective analysis, enhancing detection accuracy and interpretability.