AI 基礎設施
AtumAI Compiles Data Center Control Policies Into Searchable Specifications, Beating Human Baselines Across Three Task Categories
AtumAI converts natural-language requirements into an intermediate representation containing objectives, hard constraints, decision variables, and evaluation methods, then combines diffusion models, evolutionary algorithms, and surrogate models to search for policies. The study reports improvements in workload placement, resource scaling, and power management, but the results come from simulators and do not yet demonstrate that the system can safely take over a production control plane.
Data center scheduling, resource scaling, and power management typically rely on experts repeatedly designing heuristic policies. Allowing an LLM to rewrite controllers directly, however, can easily result in hard constraints involving capacity, latency, and power being overlooked. Submitted on August 3, AtumAI divides this work into two layers—“problem compilation” and “policy search”—with the goal of transforming the agent from a code generator into a specification-constrained system optimization tool.
The front-end Datacenter Task Compiler first uses natural-language requirements and data such as workload traces and platform constraints to construct a typed intermediate representation. Decision variables, optimization objectives, constraints, simulators, metrics, and execution-time budgets must all be explicitly specified. A rule-based critic rejects undefined variables, unit conflicts, and constraints that cannot be measured, preventing the model from inventing values. The system also maintains reusable optimization passes that can be applied across tasks, such as capacity protection and burst-load forecasting, selecting and specializing them according to the obligations of each new task.
The back end does more than ask an LLM to propose policies: a diffusion model explores control policies with different structures, an evolutionary algorithm tunes continuous or discrete parameters, and surrogate models first filter out low-potential candidates before sending a small number of options to a high-fidelity simulator. According to the paper’s reported results, workload placement increased the success rate by 17% and scheduling throughput by 8% relative to a human-designed baseline. Resource scaling improved cost efficiency by 24% while maintaining an SLO violation rate of 1.3%. Power management reduced power consumption by 21% while increasing throughput by 17%.
The engineering value lies in separating “the agent proposes a policy” from “the policy satisfies an executable contract,” while allowing the same search pipeline to serve different control problems. However, the paper does not release source code, and its evaluation is based primarily on three research environments and simulators configured by the team. Simulator bias, failure modes, policy rollback, and the stability of online changes have yet to be validated. The next things to watch are whether the intermediate representation, workloads, and baseline implementations are released, and whether candidate policies can pass shadow deployments, formal safety boundaries, and long-term testing on real clusters.