AI agents are commonly aligned with "human values" through reinforcement learning from human feedback (RLHF), where a single reward model is learned from aggregated human feedback and used to align an agent's behavior. However, human values are not homogeneous--different people hold distinct and sometimes conflicting values. Aggregating feedback into a single reward model risks disproportionately suppressing minority preferences. To address this, we present a novel reward modeling approach for learning individualized reward models. Our approach uses a language model to guide users through reflective dialogues where they critique agent behavior and construct their preferences. This personalized dialogue history, containing the user's reflections and critiqued examples, is then used as context for another language model that serves as an individualized reward function (what we call a "verbal reward model") for evaluating new trajectories. In studies with 30 participants, our method achieved a 9-12% improvement in accuracy over non-reflective verbal reward models while being more sample efficient than traditional supervised learning methods.

Butterflies, or 4-cycles in bipartite graphs, are crucial for identifying cohesive structures and dense subgraphs. While agent-based data mining is gaining prominence, its application to bipartite networks remains relatively unexplored. We propose distributed, agent-based algorithms for \emph{Butterfly Counting} in a bipartite graph $G((A,B),E)$. Agents first determine their respective partitions and collaboratively construct a spanning tree, electing a leader within $O(n \log λ)$ rounds using only $O(\log λ)$ bits per agent. A novel meeting mechanism between adjacent agents improves efficiency and eliminates the need for prior knowledge of the graph, requiring only the highest agent ID $λ$ among the $n$ agents. Notably, our techniques naturally extend to general graphs, where leader election and spanning tree construction maintain the same round and memory complexities. Building on these foundations, agents count butterflies per node in $O(Δ)$ rounds and compute the total butterfly count of $G$ in $O(Δ+\min\{|A|,|B|\})$ rounds.

Active event perception, the ability to dynamically detect, track, and summarize events in real time, is essential for embodied intelligence in tasks such as human-AI collaboration, assistive robotics, and autonomous navigation. However, existing approaches often depend on predefined action spaces, annotated datasets, and extrinsic rewards, limiting their adaptability and scalability in dynamic, real-world scenarios. Inspired by cognitive theories of event perception and predictive coding, we propose EASE, a self-supervised framework that unifies spatiotemporal representation learning and embodied control through free energy minimization. EASE leverages prediction errors and entropy as intrinsic signals to segment events, summarize observations, and actively track salient actors, operating without explicit annotations or external rewards. By coupling a generative perception model with an action-driven control policy, EASE dynamically aligns predictions with observations, enabling emergent behaviors such as implicit memory, target continuity, and adaptability to novel environments. Extensive evaluations in simulation and real-world settings demonstrate EASE's ability to achieve privacy-preserving and scalable event perception, providing a robust foundation for embodied systems in unscripted, dynamic tasks.

Warning: This paper contains content that may be inappropriate or offensive. AI agents have gained significant recent attention due to their autonomous tool usage capabilities and their integration in various real-world applications. This autonomy poses novel challenges for the safety of such systems, both in single- and multi-agent scenarios. We argue that existing red teaming or safety evaluation frameworks fall short in evaluating safety risks in complex behaviors, thought processes and actions taken by agents. Moreover, they fail to consider risks in multi-agent setups where various vulnerabilities can be exposed when agents engage in complex behaviors and interactions with each other. To address this shortcoming, we introduce the term kaleidoscopic teaming which seeks to capture complex and wide range of vulnerabilities that can happen in agents both in single-agent and multi-agent scenarios. We also present a new kaleidoscopic teaming framework that generates a diverse array of scenarios modeling real-world human societies. Our framework evaluates safety of agents in both single-agent and multi-agent setups. In single-agent setup, an agent is given a scenario that it needs to complete using the tools it has access to. In multi-agent setup, multiple agents either compete against or cooperate together to complete a task in the scenario through which we capture existing safety vulnerabilities in agents. We introduce new in-context optimization techniques that can be used in our kaleidoscopic teaming framework to generate better scenarios for safety analysis. Lastly, we present appropriate metrics that can be used along with our framework to measure safety of agents. Utilizing our kaleidoscopic teaming framework, we identify vulnerabilities in various models with respect to their safety in agentic use-cases.

Supply chain operations generate vast amounts of operational data; however, critical knowledge such as system usage practices, troubleshooting workflows, and resolution techniques often remains buried within unstructured communications like support tickets, emails, and chat logs. While RAG systems aim to leverage such communications as a knowledge base, their effectiveness is limited by raw data challenges: support tickets are typically noisy, inconsistent, and incomplete, making direct retrieval suboptimal. Unlike existing RAG approaches that focus on runtime optimization, we introduce a novel offline-first methodology that transforms these communications into a structured knowledge base. Our key innovation is a LLMs-based multi-agent system orchestrating three specialized agents: Category Discovery for taxonomy creation, Categorization for ticket grouping, and Knowledge Synthesis for article generation. Applying our methodology to real-world support tickets with resolution notes and comments, our system creates a compact knowledge base - reducing total volume to just 3.4% of original ticket data while improving quality. Experiments demonstrate that our prebuilt knowledge base in RAG systems significantly outperforms traditional RAG implementations (48.74% vs. 38.60% helpful answers) and achieves a 77.4% reduction in unhelpful responses. By automating institutional knowledge capture that typically remains siloed in experts' heads, our solution translates to substantial operational efficiency: reducing support workload, accelerating resolution times, and creating self-improving systems that automatically resolve approximately 50% of future supply chain tickets. Our approach addresses a key gap in knowledge management by transforming transient communications into structured, reusable knowledge through intelligent offline processing rather than latency-inducing runtime architectures.

A long-term goal of Artificial Intelligence is to build a language understanding system that allows a human to collaborate with a physical robot using language that is natural to the human. In this paper we highlight some of the challenges in doing this, and propose a solution that integrates the abilities of a cognitive agent capable of interactive task learning in a physical robot with the linguistic abilities of a large language model. We also point the way to an initial implementation of this approach.

Fairness is desirable yet challenging to achieve within multi-agent systems, especially when agents differ in latent traits that affect their abilities. This hidden heterogeneity often leads to unequal distributions of wealth, even when agents operate under the same rules. Motivated by real-world examples, we propose a framework based on repeated principal-agent games, where a principal, who also can be seen as a player of the game, learns to offer adaptive contracts to agents. By leveraging a simple yet powerful contract structure, we show that a fairness-aware principal can learn homogeneous linear contracts that equalize outcomes across agents in a sequential social dilemma. Importantly, this fairness does not come at the cost of efficiency: our results demonstrate that it is possible to promote equity and stability in the system while preserving overall performance.

We contribute the first provable guarantees of global convergence to Nash equilibria (NE) in two-player zero-sum convex Markov games (cMGs) by using independent policy gradient methods. Convex Markov games, recently defined by Gemp et al. (2024), extend Markov decision processes to multi-agent settings with preferences that are convex over occupancy measures, offering a broad framework for modeling generic strategic interactions. However, even the fundamental min-max case of cMGs presents significant challenges, including inherent nonconvexity, the absence of Bellman consistency, and the complexity of the infinite horizon. We follow a two-step approach. First, leveraging properties of hidden-convex--hidden-concave functions, we show that a simple nonconvex regularization transforms the min-max optimization problem into a nonconvex-proximal Polyak-Lojasiewicz (NC-pPL) objective. Crucially, this regularization can stabilize the iterates of independent policy gradient methods and ultimately lead them to converge to equilibria. Second, building on this reduction, we address the general constrained min-max problems under NC-pPL and two-sided pPL conditions, providing the first global convergence guarantees for stochastic nested and alternating gradient descent-ascent methods, which we believe may be of independent interest.

Large Language Models (LLMs) have shown strong capability in diverse software engineering tasks, e.g. code completion, bug fixing, and document generation. However, feature-driven development (FDD), a highly prevalent real-world task that involves developing new functionalities for large, existing codebases, remains underexplored. We therefore introduce SWE-Dev, the first large-scale dataset (with 14,000 training and 500 test samples) designed to evaluate and train autonomous coding systems on real-world feature development tasks. To ensure verifiable and diverse training, SWE-Dev uniquely provides all instances with a runnable environment and its developer-authored executable unit tests. This collection not only provides high-quality data for Supervised Fine-Tuning (SFT), but also enables Reinforcement Learning (RL) by delivering accurate reward signals from executable unit tests. Our extensive evaluations on SWE-Dev, covering 17 chatbot LLMs, 10 reasoning models, and 10 Multi-Agent Systems (MAS), reveal that FDD is a profoundly challenging frontier for current AI (e.g., Claude-3.7-Sonnet achieves only 22.45\% Pass@3 on the hard test split). Crucially, we demonstrate that SWE-Dev serves as an effective platform for model improvement: fine-tuning on training set enabled a 7B model comparable to GPT-4o on \textit{hard} split, underscoring the value of its high-quality training data. Code is available here \href{https://github.com/DorothyDUUU/SWE-Dev}{https://github.com/DorothyDUUU/SWE-Dev}.

As AI capabilities advance toward and potentially beyond human-level performance, a natural transition emerges where AI-driven development becomes more efficient than human-centric approaches. A promising pathway toward this transition lies in AI-for-AI (AI4AI), which leverages AI techniques to automate and optimize the design, training, and deployment of AI systems themselves. While LLM-based agents have shown the potential to realize AI4AI, they are often unable to fully leverage the experience accumulated by agents during the exploration of solutions in the reasoning process, leading to inefficiencies and suboptimal performance. To address this limitation, we propose ML-Master, a novel AI4AI agent that seamlessly integrates exploration and reasoning by employing a selectively scoped memory mechanism. This approach allows ML-Master to efficiently combine diverse insights from parallel solution trajectories with analytical reasoning, guiding further exploration without overwhelming the agent with excessive context. We evaluate ML-Master on the MLE-Bench, where it achieves a 29.3% average medal rate, significantly surpassing existing methods, particularly in medium-complexity tasks, while accomplishing this superior performance within a strict 12-hour time constraint-half the 24-hour limit used by previous baselines. These results demonstrate ML-Master's potential as a powerful tool for advancing AI4AI.