Optimizing recommender systems for objectives beyond accuracy, such as diversity, novelty, and personalization, is crucial for long-term user satisfaction. To this end, industrial practitioners have accumulated vast amounts of structured domain knowledge, which we term human priors (e.g., item taxonomies, temporal patterns). This knowledge is typically applied through post-hoc adjustments during ranking or post-ranking. However, this approach remains decoupled from the core model learning, which is particularly undesirable as the industry shifts to end-to-end generative recommendation foundation models. On the other hand, many methods targeting these beyond-accuracy objectives often require architecture-specific modifications and discard these valuable human priors by learning user intent in a fully unsupervised manner. Instead of discarding the human priors accumulated over years of practice, we introduce a backbone-agnostic framework that seamlessly integrates these human priors directly into the end-to-end training of generative recommenders. With lightweight, prior-conditioned adapter heads inspired by efficient LLM decoding strategies, our approach guides the model to disentangle user intent along human-understandable axes (e.g., interaction types, long- vs. short-term interests). We also introduce a hierarchical composition strategy for modeling complex interactions across different prior types. Extensive experiments on three large-scale datasets demonstrate that our method significantly enhances both accuracy and beyond-accuracy objectives. We also show that human priors allow the backbone model to more effectively leverage longer context lengths and larger model sizes.

Managing a vast number of virtual machines (VMs) efficiently is a critical challenge in modern large-scale data centers. The continuous creation and termination of VMs lead to resource fragmentation across physical machines (PMs), necessitating periodic VM rescheduling to optimize resource utilization. Despite its significance, VM rescheduling has received limited attention in the literature. A key challenge is that, unlike conventional combinatorial optimization problems, the efficiency of rescheduling algorithms is heavily impacted by inference time, as VM states evolve dynamically during execution. This scalability bottleneck hampers existing methods. To address this, we propose VMR$^{2}$L, a reinforcement learning framework tailored for VM rescheduling. VMR$^{2}$L integrates a two-stage decision-making process to accommodate complex operational constraints, a feature extraction mechanism that captures critical relational information for rescheduling, and a risk-aware evaluation strategy that enables users to balance execution speed and rescheduling accuracy. Extensive experiments using real-world data from a production-scale data center demonstrate that VMR$^{2}$L achieves near-optimal performance while reducing inference time to a matter of seconds. To facilitate reproducibility, we provide access to our implementation and datasets.

Modern industry-scale data centers need to manage a large number of virtual machines (VMs). Due to the continual creation and release of VMs, many small resource fragments are scattered across physical machines (PMs). To handle these fragments, data centers periodically reschedule some VMs to alternative PMs, a practice commonly referred to as VM rescheduling. Despite the increasing importance of VM rescheduling as data centers grow in size, the problem remains understudied. We first show that, unlike most combinatorial optimization tasks, the inference time of VM rescheduling algorithms significantly influences their performance, due to dynamic VM state changes during this period. This causes existing methods to scale poorly. Therefore, we develop a reinforcement learning system for VM rescheduling, VMR2L, which incorporates a set of customized techniques, such as a two-stage framework that accommodates diverse constraints and workload conditions, a feature extraction module that captures relational information specific to rescheduling, as well as a risk-seeking evaluation enabling users to optimize the trade-off between latency and accuracy. We conduct extensive experiments with data from an industry-scale data center. Our results show that VMR2L can achieve a performance comparable to the optimal solution but with a running time of seconds. Code12 and datasets3 are open-sourced.

Containerization has garnered substantial favor among cloud service providers. Nevertheless, the notable network overhead incurred between containers has prompted concerns within the community. In cloud resource scheduling, collocating service containers that frequently communicate to the same machine - termed “service affinity” - is instrumental in enhancing application performance. In response to this concern, we present a solution that harnesses service affinity and collocates containers to enhance the overall system performance and stability. To maximize the benefits of collocating containers, it is necessary to calculate a new schedule that optimally and efficiently maximizes service affinity, especially within the expansive domain of industry-scale cloud environments. In pursuit of this, we leverage the skewness property of affinity and machine learning to fuse solver-based algorithms, thereby assuring both quality and efficiency for problems at scale. Our methodology encompasses the partitioning of a given task into discrete subproblems, with a keen focus on resolving the most critical ones. Via a graph neural network classifier, we assign each subproblem to be solved independently using methods based on off-the-shelf solvers in our algorithm pool - namely, MIP-based, or column generation. This strategic approach enables the efficient computation of a schedule for a cloud cluster that fully optimizes the overall service affinity. We further propose a heuristic algorithm to compute executable container migration plans for practical use, facilitating the transition to the new placement where service affinity is well optimized. Our solution has been deployed in our large-scale production environment, covering over a million cores within ByteDance. Through the successful real-world production deployment, our approach exhibits an average improvement in end-to-end latency by 23.75% and a reduction in request error rates by 24.09% compared to the original system.

We investigate safe multi-agent reinforcement learning, where agents seek to collectively maximize an aggregate sum of local objectives while satisfying their own safety constraints. The objective and constraints are described by general utilities, i.e., nonlinear functions of the long-term state-action occupancy measure, which encompass broader decision-making goals such as risk, exploration, or imitations. The exponential growth of the state-action space size with the number of agents presents challenges for global observability, further exacerbated by the global coupling arising from agents' safety constraints. To tackle this issue, we propose a primal-dual method utilizing shadow reward and $\kappa$-hop neighbor truncation under a form of correlation decay property, where $\kappa$ is the communication radius. In the exact setting, our algorithm converges to a first-order stationary point (FOSP) at the rate of $\mathcal{O}\left(T^{-2/3}\right)$. In the sample-based setting, we demonstrate that, with high probability, our algorithm requires $\widetilde{\mathcal{O}}\left(\epsilon^{-3.5}\right)$ samples to achieve an $\epsilon$-FOSP with an approximation error of $\mathcal{O}(\phi_0^{2\kappa})$, where $\phi_0\in (0,1)$. Finally, we demonstrate the effectiveness of our model through extensive numerical experiments.