AI research agents are demonstrating great potential to accelerate scientific progress by automating the design, implementation, and training of machine learning models. We focus on methods for improving agents' performance on MLE-bench, a challenging benchmark where agents compete in Kaggle competitions to solve real-world machine learning problems. We formalize AI research agents as search policies that navigate a space of candidate solutions, iteratively modifying them using operators. By designing and systematically varying different operator sets and search policies (Greedy, MCTS, Evolutionary), we show that their interplay is critical for achieving high performance. Our best pairing of search strategy and operator set achieves a state-of-the-art result on MLE-bench lite, increasing the success rate of achieving a Kaggle medal from 39.6% to 47.7%. Our investigation underscores the importance of jointly considering the search strategy, operator design, and evaluation methodology in advancing automated machine learning.

As AI becomes more personal, e.g., Agentic AI, there is an increasing need for personalizing models for various use cases. Personalized federated learning (PFL) enables each client to collaboratively leverage other clients' knowledge for better adaptation to the task of interest, without privacy risks. Despite its potential, existing PFL methods remain confined to rather simplified scenarios where data and models are the same across clients. To move towards realistic scenarios, we propose FedMosaic, a method that jointly addresses data and model heterogeneity with a task-relevance-aware model aggregation strategy to reduce parameter interference, and a dimension-invariant module that enables knowledge sharing across heterogeneous architectures without huge computational cost. To mimic the real-world task diversity, we propose a multi-modal PFL benchmark spanning 40 distinct tasks with distribution shifts over time. The empirical study shows that FedMosaic outperforms the state-of-the-art PFL methods, excelling in both personalization and generalization capabilities under challenging, realistic scenarios.

Discovering the antecedents of individuals' influence in collaborative environments is an important, practical, and challenging problem. In this paper, we study interpersonal influence in small groups of individuals who collectively execute a sequence of intellective tasks. We observe that along an issue sequence with feedback, individuals with higher expertise and social confidence are accorded higher interpersonal influence. We also observe that low-performing individuals tend to underestimate their high-performing teammate's expertise. Based on these observations, we introduce three hypotheses and present empirical and theoretical support for their validity. We report empirical evidence on longstanding theories of transactive memory systems, social comparison, and confidence heuristics on the origins of social influence. We propose a cognitive dynamical model inspired by these theories to describe the process by which individuals adjust interpersonal influences over time. We demonstrate the model's accuracy in predicting individuals' influence and provide analytical results on its asymptotic behavior for the case with identically performing individuals. Lastly, we propose a novel approach using deep neural networks on a pre-trained text embedding model for predicting the influence of individuals. Using message contents, message times, and individual correctness collected during tasks, we are able to accurately predict individuals' self-reported influence over time. Extensive experiments verify the accuracy of the proposed models compared to baselines such as structural balance and reflected appraisal model. While the neural networks model is the most accurate, the dynamical model is the most interpretable for influence prediction.

Convolutional neural networks (CNNs) underpin many modern computer vision systems. With applications ranging from common to critical areas, a need to explain and understand the model and its decisions (XAI) emerged. Prior works suggest that in the top layers of CNNs, the individual channels can be attributed to classifying human-understandable concepts. Concept relevance propagation (CRP) methods can backtrack predictions to these channels and find images that most activate these channels. However, current CRP workflows are largely manual: experts must inspect activation images to name the discovered concepts and must synthesize verbose explanations from relevance maps, limiting the accessibility of the explanations and their scalability. To address these issues, we introduce Large Language model EXplaIns COncept Relevance Propagation (LLEXICORP), a modular pipeline that couples CRP with a multimodal large language model. Our approach automatically assigns descriptive names to concept prototypes and generates natural-language explanations that translate quantitative relevance distributions into intuitive narratives. To ensure faithfulness, we craft prompts that teach the language model the semantics of CRP through examples and enforce a separation between naming and explanation tasks. The resulting text can be tailored to different audiences, offering low-level technical descriptions for experts and high-level summaries for non-technical stakeholders. We qualitatively evaluate our method on various images from ImageNet on a VGG16 model. Our findings suggest that integrating concept-based attribution methods with large language models can significantly lower the barrier to interpreting deep neural networks, paving the way for more transparent AI systems.

Knowledge tracing (KT) models are a crucial basis for pedagogical decision-making, namely which task to select next for a learner and when to stop teaching a particular skill. Given the high stakes of pedagogical decisions, KT models are typically required to be interpretable, in the sense that they should implement an explicit model of human learning and provide explicit estimates of learners' abilities. However, to our knowledge, no study to date has investigated whether the interpretability of KT models actually helps human teachers to make teaching decisions. We address this gap. First, we perform a simulation study to show that, indeed, decisions based on interpretable KT models achieve mastery faster compared to decisions based on a non-interpretable model. Second, we repeat the study but ask $N=12$ human teachers to make the teaching decisions based on the information provided by KT models. As expected, teachers rate interpretable KT models higher in terms of usability and trustworthiness. However, the number of tasks needed until mastery hardly differs between KT models. This suggests that the relationship between model interpretability and teacher decisions is not straightforward: teachers do not solely rely on KT models to make decisions and further research is needed to investigate how learners and teachers actually understand and use KT models.

Training agents to operate under strict constraints during deployment, such as limited resource budgets or stringent safety requirements, presents significant challenges, especially when these constraints render the task complex. In this work, we propose a curriculum learning strategy that gradually tightens constraints during training, enabling the agent to incrementally master the deployment requirements. Inspired by self-paced learning techniques in unconstrained reinforcement learning (RL), our approach facilitates a smoother transition to challenging environments by initially training on simplified versions of the constraints and progressively introducing the full deployment conditions. We provide a theoretical analysis using an RL agent in a binary-tree Markov Decision Process (MDP) to demonstrate that our curriculum strategy can accelerate training relative to a baseline approach that imposes the trajectory constraints from the outset. Moreover, we empirically validate the effectiveness and generality of our method across both RL and large language model (LLM) agents in diverse settings, including a binary-tree MDP, a multi-task navigation domain, and a math reasoning task with two benchmarks. These results highlight the potential of curriculum design in enhancing the efficiency and performance of agents operating under complex trajectory constraints during deployment. Moreover, when applied to LLMs, our strategy enables compression of output chain-of-thought tokens, achieving a substantial inference speedup on consumer hardware, demonstrating its effectiveness for resource-constrained deployment.

The trajectory of AI development suggests that we will increasingly rely on agent-based systems composed of independently developed agents with different information, privileges, and tools. The success of these systems will critically depend on effective collaboration among these heterogeneous agents, even under partial observability. Despite intense interest, few empirical studies have evaluated such agent-agent collaboration at scale. We propose a collaborative maze-solving benchmark that (i) isolates collaborative capabilities, (ii) modulates problem complexity, (iii) enables scalable automated grading, and (iv) imposes no output-format constraints, preserving ecological plausibility. Using this framework, we evaluate 32 leading open- and closed-source models in solo, homogeneous, and heterogeneous pairings. Our results reveal a "collaboration gap": models that perform well solo often degrade substantially when required to collaborate. Collaboration can break down dramatically; for instance, small distilled models that solve mazes well alone may fail almost completely in certain pairings. We find that starting with the stronger agent often improves outcomes, motivating a "relay inference" approach where the stronger agent leads before handing off to the weaker one, closing much of the gap. Our findings argue for (1) collaboration-aware evaluation, (2) training strategies developed to enhance collaborative capabilities, and (3) interaction design that reliably elicits agents' latent skills, guidance that applies to AI-AI and human-AI collaboration.

Training and education in human-centered fields require authentic practice, yet realistic simulations of human behavior have remained limited. We present a multi-agent psychological simulation system that models internal cognitive-affective processes to generate believable human behaviors. In contrast to black-box neural models, this system is grounded in established psychological theories (e.g., self-efficacy, mindset, social constructivism) and explicitly simulates an ``inner parliament'' of agents corresponding to key psychological factors. These agents deliberate and interact to determine the system's output behavior, enabling unprecedented transparency and alignment with human psychology. We describe the system's architecture and theoretical foundations, illustrate its use in teacher training and research, and discuss how it embodies principles of social learning, cognitive apprenticeship, deliberate practice, and meta-cognition.

We study a many-to-one matching problem, such as the college admission problem, where each college can admit multiple students. Unlike classical models, colleges evaluate sets of students through non-linear utility functions that capture diversity between them. In this setting, we show that classical stable matchings may fail to exist. To address this, we propose alternative solution concepts based on Rawlsian fairness, aiming to maximize the minimum utility across colleges. We design both deterministic and stochastic algorithms that iteratively improve the outcome of the worst-off college, offering a practical approach to fair allocation when stability cannot be guaranteed.

This work has been submitted to the IEEE for possible publication. Abstract--Endowing robot hands with human-level dexterity has been a long-standing goal in robotics. Bimanual robotic piano playing represents a particularly challenging task: it is high-dimensional, contact-rich, and requires fast, precise control. Our approach is built on three core components. First, we introduce an automatic fingering strategy based on Optimal Transport (OT), allowing the agent to autonomously discover efficient piano-playing strategies from scratch without demonstrations. Second, we conduct large-scale Reinforcement Learning (RL) by training more than 2,000 agents, each specialized in distinct music pieces, and aggregate their experience into a dataset named RP1M++, consisting of over one million trajectories for robotic piano playing. Extensive experiments and ablation studies highlight the effectiveness and scalability of our approach, advancing dexterous robotic piano playing at scale. Achieving human-level dexterity remains one of the central challenges in robotics. The difficulty stems from the breadth of challenges ranging from contact-rich manipulation to dynamic athletic tasks, each posing distinct demands. Manipulation tasks, such as grasping or reorienting objects [1], require sustained application of appropriate forces at moderate speeds across objects with diverse shapes, materials, and weight distributions. Dynamic tasks, such as juggling [2] or table tennis [3], involve frequent contact changes, demand high precision, and allow little tolerance for error due to the rarity of contact opportunities. The combination of requiring both precision and speed makes reproducing human-level dexterity particularly challenging. Q. Gao is with the University of Southern California, CA 90007, United States (e-mail: quankaig@usc.edu). Q. Cheng is with Imperial College London, SW7 2AZ, London, United Kingdom (e-mail: c.qian24@imperial.ac.uk). J. Kannala is with the University of Oulu, 90570 Oulu, Finland. D. B uchler is also with the University of Alberta (Canada), the Alberta Machine Intelligence Institute (Amii), & holds a Canada CIFAR AI Chair.