manage population diversity

This study investigates how behavioral diversity influences team performance in multi-agent reinforcement learning (MARL), particularly under computational constraints and sparse reward settings, to enhance collaboration efficiency and robustness. We propose a trajectory-embedding-based behavioral distance metric, a diversity-regularized objective, heterogeneous policy initialization, and a curriculum-style perturbation training framework. Our work is the first to systematically demonstrate that behavioral heterogeneity spontaneously induces unbiased role specialization, strengthens morphological synergy, accelerates cooperative policy discovery under sparse rewards, and enables implicit skill retention and transfer. Experiments across diverse collaborative tasks show that heterogeneous teams achieve 23–41% higher average task success rates than homogeneous baselines, recover from environmental perturbations 2.8× faster, and more stably acquire reusable collaborative sub-policies.

In cooperative multi-agent reinforcement learning, the theoretical conditions under which heterogeneous teams outperform homogeneous ones remain poorly understood—particularly in task allocation settings. Method: This paper investigates how behavioral diversity enhances team performance through the lens of reward design, introducing a curvature-based theoretical criterion for reward functions to identify sufficient conditions for heterogeneity-induced gains. It further proposes Gradient-Driven Heterogeneous Environment Design (HED), a differentiable algorithm that constructs tasks explicitly amplifying diversity advantages. Contribution/Results: Leveraging generalized aggregation analysis, MARL modeling, and differentiable environment optimization, the work empirically validates—across matrix games and embodied multi-objective capture tasks—that convex reward structures maximally benefit heterogeneous teams, significantly surpassing homogeneous baselines. The core contribution is the first formal linkage between reward function curvature and heterogeneity advantage, establishing an automated paradigm for co-optimizing environment, reward, and agent architecture.

Unsupervised skill discovery faces a fundamental trade-off between behavioral quality and diversity—particularly challenging in high-dimensional robotic control domains with rich latent skill spaces. To address this, we propose Contrastive Multi-objective Skill Discovery (ComSD), the first framework to jointly optimize a contrastive learning–driven diversity reward and a particle-filter–based exploration reward, thereby establishing a dynamic multi-objective reward mechanism in a reward-free setting. ComSD integrates adaptive weight balancing and unsupervised representation learning to unify high-fidelity skill discrimination with progressive discovery of novel behaviors. Evaluated on 32 downstream tasks, ComSD achieves state-of-the-art performance, significantly enhancing both behavioral diversity and practical utility of multi-joint robots across hierarchical exploration regimes.

This work addresses the collapse of generation diversity in reinforcement fine-tuning, where optimization dynamics often drive model outputs toward a single solution (i.e., a Dirac delta distribution) due to misalignment between the objective function and the optimization landscape. To mitigate this, we propose DRIFT, the first framework to systematically incorporate diversity incentives into reinforcement fine-tuning. DRIFT synergistically preserves both task alignment and output diversity during policy updates through reward-concentrated subset sampling, stochastic prompt augmentation, and potential-based reward shaping. Experimental results demonstrate that DRIFT achieves Pareto superiority: it improves generation diversity by 9.08%–43.46% while maintaining equivalent task alignment, or enhances task alignment by 59.65%–65.86% under comparable diversity levels.

Pretrained generative models for drug molecule design often suffer from premature convergence to local optima during reinforcement learning (RL)-based reward optimization, resulting in limited molecular diversity and suboptimal drug-likeness. To address this, we propose an RL framework featuring adaptive reward function updating. We systematically investigate diverse intrinsic motivation mechanisms for controlling molecular diversity and introduce a novel synergistic reward correction strategy that jointly incorporates structural similarity penalization and uncertainty-aware predictive rewards. Our method integrates graph neural networks (GNNs) with policy gradient optimization. Evaluated on multiple benchmark datasets, the generated molecule sets achieve a 37% average improvement in diversity—measured by scaffold and fingerprint dissimilarity—while maintaining or improving drug-likeness (quantitative estimate of drug-likeness, QED; synthetic accessibility, SA) and target-binding activity (pIC₅₀). The framework significantly outperforms state-of-the-art baselines, demonstrating superior balance between exploration and exploitation in de novo molecular generation.

This study addresses the persistent issue of diversity collapse in multi-agent large language models during open-ended creative generation, which severely constrains collective exploration yet remains mechanistically unclear. Adopting a structural coupling perspective, the work systematically investigates how interaction architectures precipitate diversity loss, demonstrating that the root cause lies in system design rather than inherent model limitations. Through multi-level controlled experiments, the authors examine the effects of model capability, role authority, group size, and communication topology, revealing that strong alignment yields diminishing diversity returns, authority dominance suppresses semantic variety, and dense communication accelerates premature convergence. The findings underscore the critical importance of preserving agent independence and constructive disagreement in creative tasks, highlighting the decisive role of structural design in sustaining diversity.

Traditional reinforcement learning relies on deterministic policies, which struggle to meet the demand for behavioral diversity in tasks such as language model fine-tuning or scientific discovery. This work proposes a novel paradigm based on distributions over reward functions, introducing nonlinear objectives over action sets and leveraging a principled gradient estimator derived from contextual bandits. The approach enables controllable induction of policy diversity without compromising expected return. By unifying classical policy gradient methods with action-set optimization frameworks, the proposed method demonstrates robust generation of diverse behaviors in complex tasks, significantly outperforming conventional approaches in empirical evaluations.

This study investigates how functional diversity influences team communication and performance across varying contexts, with particular attention to the mechanisms at play when teams lack specific expertise. By constructing a multi-agent simulation model, the research systematically examines the joint effects of individual functional diversity (IFD), dominant functional diversity (DFD), and aggregated team-level professional capability under heterogeneous functional compositions and diverse communication structures. The findings reveal that the impact of functional diversity on team performance is highly contingent on both communication structure and functional composition—enhancing performance in some contexts while impairing it in others. This challenges conventional approaches that focus solely on lower-order moments of capability distributions and underscores the necessity and explanatory power of incorporating a holistic measure of team-level professional capacity.

This study addresses the critical gap in understanding how team diversity influences fairness in AI systems, which are often exacerbated by data and design flaws that perpetuate social inequities. Drawing on grounded theory, the authors conducted in-depth interviews with 25 practitioners across four AI development teams at a major software company in Brazil and Portugal, working on projects spanning education, energy, accessibility, and facial recognition. The research systematically identifies six key roles that social diversity plays in AI development: recognizing bias, infusing empathy, confronting systemic discrimination, fostering inclusive decision-making, serving as a safeguard against bias, and broadening problem-solving perspectives. Findings demonstrate that diverse teams significantly enhance the fairness and inclusivity of AI systems and offer actionable pathways for integrating fairness into software engineering practices.

This work addresses the issue of premature convergence in multi-agent multi-objective optimization, which often arises from behavioral homogenization. To mitigate this, the study introduces a behavioral entropy maximization mechanism into multi-objective evolutionary algorithms for the first time. Specifically, within the NSGA-II framework, it integrates policy entropy rewards with multi-objective fitness evaluation to explicitly promote behavioral diversity while preserving Pareto optimality. This approach effectively alleviates behavioral collapse and substantially enhances exploration capability. Experimental results in the rover domain demonstrate that, compared to the NSGA-II baseline, the proposed method achieves up to a 48% improvement in hypervolume metric, along with significantly enhanced solution set quality and diversity.