- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
This paper proposes InfraMind, a multi-agent orchestration framework that uses serving load and time budgets when deciding how much reasoning to perform. A planner selects the collaboration graph when a request arrives, and an executor chooses the model and reasoning prompt before each agent call using refreshed serving telemetry. The two policies share a latency penalty during training, and pending calls are scheduled by deadline. Experiments on five benchmarks measure accuracy, latency, and correct answers delivered on time. In the 15-model study with replayed Azure arrival timings, the paper reports a 4.7x improvement in correct, on-time answers over MasRouter.
Critical strengths
-
The paper identifies a real limitation of task-only orchestration: a model or reasoning strategy that is useful in isolation can be a poor choice when its queue is congested. Adapting the requested computation to the available serving capacity is a useful direction. Measuring correct, on-time answers also captures the practical objective better than reporting accuracy and latency separately.
-
The common-EDF experiment on MATH provides evidence that orchestration changes have value beyond changing the queue scheduler. Delivered accuracy improves from 20.4% for MasRouter to 33.8% for InfraMind under this control. The executor telemetry ablation is also helpful for understanding why observing the serving state matters.
Critical weaknesses
-
The comparisons do not consistently match scheduling and training conditions. In Table 1, InfraMind uses EDF on MATH, HumanEval, and MMLU-Pro, while the baselines use FCFS. The common-EDF MATH experiment shows that the benefit persists there, but does not resolve the scheduling contribution on the other two tasks. In the 15-model study, Appendix D reports 15 training epochs for InfraMind versus five for MasRouter, additional load and budget training for InfraMind, and a zero cost-quality coefficient for MasRouter. These differences may favor InfraMind and make the size of the policy's advantage difficult to judge.
-
The added value of the planner and shared penalty remains unclear. The telemetry ablation supports using serving-state inputs, but does not establish why the full hierarchy is needed over a simpler load-aware extension of MasRouter. Section 5.3 evaluates load-conditioned planning, the shared penalty, and joint model-strategy selection together. The discussion would benefit from less repeated motivation and a clearer explanation of what these choices contribute individually.
The writing in this paper omits context needed for human readers, while creating unnecessarily long, repetitive basic claims and presenting long lists of experimental numbers with little interpretation.
Questions for authors
-
How do the HumanEval and MMLU-Pro comparisons change when all methods use the same queue scheduler, as in the common-EDF MATH experiment?
-
In the 15-model experiment, does the advantage persist when MasRouter receives comparable training effort and an objective suited to the same deadlines?
-
How much does the load-conditioned planner contribute when the executor is already load-aware? Would the original task-driven planner paired with the proposed executor achieve similar results?