Industrial question-answering (QA) systems require higher safety and reliability than general-purpose dialogue models, as errors in high-risk scenarios such as equipment fault diagnosis can have severe consequences. Although multi-agent large language models enhance reasoning depth, they suffer from uncontrolled iterations and unverifiable outputs, and conventional distillation methods struggle to transfer collaborative reasoning capabilities to lightweight, deployable student models. To address these challenges, we propose Knowledge Graph-guided Multi-Agent System Distillation (KG-MASD). Our approach formulates distillation as a Markov Decision Process and incorporates a knowledge graph as a verifiable structured prior to enrich state representation and ensure convergence. By integrating collaborative reasoning with knowledge grounding, KG-MASD generates high-confidence instruction-tuning data and jointly distills reasoning depth and verifiability into compact student models suitable for edge deployment. Experiments on an industrial QA dataset show that KG-MASD improves accuracy by 2.4 per cent to 20.1 per cent over baselines and significantly enhances reliability, enabling trustworthy AI deployment in safety-critical industrial scenarios. Code and data are available at https://github.com/erwinmsmith/KG-MAD/.

Multi-agent social interaction has clearly benefited from Large Language Models. However, current simulation systems still face challenges such as difficulties in scaling to diverse scenarios and poor reusability due to a lack of modular design. To address these issues, we designed and developed a modular, object-oriented framework that organically integrates various base classes through a hierarchical structure, harvesting scalability and reusability. We inherited the framework to realize common derived classes. Additionally, a memory summarization mechanism is proposed to filter and distill relevant information from raw memory data, prioritizing contextually salient events and interactions. By selecting and combining some necessary derived classes, we customized a specific simulated environment. Utilizing this simulated environment, we successfully simulated human interactions on social media, replicating real-world online social behaviors. The source code for the project will be released and evolve.

With recent advancements in multi-agent generative AI (Gen AI), technology organizations like Microsoft are adopting these complex tools, redefining AI agents as active collaborators in complex workflows rather than as passive tools. In this study, we investigated how early adopters and developers conceptualize multi-agent Gen AI tools, focusing on how they understand human-AI collaboration mechanisms, general collaboration dynamics, and transparency in the context of AI tools. We conducted semi-structured interviews with 13 developers, all early adopters of multi-agent Gen AI technology who work at Microsoft. Our findings revealed that these early adopters conceptualize multi-agent systems as "teams" of specialized role-based and task-based agents, such as assistants or reviewers, structured similar to human collaboration models and ranging from AI-dominant to AI-assisted, user-controlled interactions. We identified key challenges, including error propagation, unpredictable and unproductive agent loop behavior, and the need for clear communication to mitigate the layered transparency issues. Early adopters' perspectives about the role of transparency underscored its importance as a way to build trust, verify and trace errors, and prevent misuse, errors, and leaks. The insights and design considerations we present contribute to CSCW research about collaborative mechanisms with capabilities ranging from AI-dominant to AI-assisted interactions, transparency and oversight strategies in human-agent and agent-agent interactions, and how humans make sense of these multi-agent systems as dynamic, role-diverse collaborators which are customizable for diverse needs and workflows. We conclude with future research directions that extend CSCW approaches to the design of inter-agent and human mediation interactions.

Social intelligence has become a critical capability for large language models (LLMs), enabling them to engage effectively in real-world social tasks such as collaboration and negotiation. Reinforcement learning (RL) is a natural fit for training socially intelligent agents because it allows models to learn sophisticated strategies directly through social interactions without requiring human annotations. However, there are two unique parts about social intelligence tasks: (1) the quality of individual utterances in social interactions is not strictly related to final success; (2) social interactions require multi-dimensional rubrics for success. Therefore, we argue that it is necessary to design rewards for building utterance-level multi-dimensional reward models to facilitate RL training for social intelligence tasks. To address these challenges, we propose Sotopia-RL, a novel framework that refines coarse episode-level feedback into utterance-level, multi-dimensional rewards. Utterance-level credit assignment attributes outcomes to individual utterances, while multi-dimensional rewards capture the full richness of social interactions and reduce reward hacking. Experiments in Sotopia, an open-ended social learning environment, demonstrate that Sotopia-RL achieves state-of-the-art social goal completion scores (7.17 on Sotopia-hard and 8.31 on Sotopia-full), significantly outperforming existing approaches. Ablation studies confirm the necessity of both utterance-level credit assignment and multi-dimensional reward design for RL training.

AI agents have significant potential to reshape cybersecurity, making a thorough assessment of their capabilities critical. However, existing evaluations fall short, because they are based on small-scale benchmarks and only measure static outcomes, failing to capture the full, dynamic range of real-world security challenges. To address these limitations, we introduce CyberGym, a large-scale benchmark featuring 1,507 real-world vulnerabilities across 188 software projects. Adjustable to different vulnerability analysis settings, CyberGym primarily tasks agents with generating a proof-of-concept test that reproduces a vulnerability, given only its text description and the corresponding codebase. Our extensive evaluation highlights that CyberGym effectively differentiates agents' and models' cybersecurity capabilities. Even the top-performing combinations only achieve a ~20% success rate, demonstrating the overall difficulty of CyberGym. Beyond static benchmarking, we show that CyberGym leads to the discovery of 35 zero-day vulnerabilities and 17 historically incomplete patches. These results underscore that CyberGym is not only a robust benchmark for measuring AI's progress in cybersecurity but also a platform for creating direct, real-world security impact.

Mechanism design is a high-impact branch of economics and computer science that studies the implementation of socially desirable outcomes among strategic self-interested agents. Major real-world use cases include combinatorial auctions ( e.g., strategic sourcing, radio spectrum auctions),

Mechanism design is a high-impact branch of economics and computer science that studies the implementation of socially desirable outcomes among strategic self-interested agents. Major real-world use cases include combinatorial auctions ( e.g., strategic sourcing, radio spectrum auctions),

In this paper, departing from the existing literature on IL for MFGs, we introduce a new solution concept called the Nash imitation gap.