Title: TIMEGATE: Sustainable Time-Boxed Promotion Gates for Continual ML Adaptation Under Resource Constraints

Abstract: The shift toward continual adaptation in machine learning (ML) systems incurs significant costs in terms of energy, computational power, and data annotation with every retraining cycle. To address this, we present TIMEGATE, a policy framework that orchestrates adaptation by strategically allocating resources across time, labeling, training, and evaluation phases. Central to this system is the metric-availability signal, M, which guides decisions on whether to conduct partial or comprehensive evaluations. Our validation studies demonstrate five key findings: (i) On the Adult tabular dataset, prioritizing labeling yields a 2.3x performance advantage over training; (ii) This approach generalizes to LLaMA-3.1-8B fine-tuned with QLoRA on SST-2, achieving an accuracy increase from 0.80 to 0.96, with the M signal equaling 1 in 35 out of 36 runs; (iii) The M signal proves highly informative, as evidenced by a 28-cell sensitivity analysis showing a drop to 0.81 under tight thresholds; (iv) A simulation spanning 100 cycles reveals that evaluation compute usage is reduced by 66%, with no undetected mis-promotions occurring; and (v) Implementing a 10%-slice evaluation strategy on LLaMA models on a single H200 GPU cuts both wall-clock time and energy consumption by 89%, with consistency ratios aligning within 0.2%.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC