A paper on distributed AI training schedules argues that most published comparisons are measuring against the wrong baseline.
Researchers studying DiLoCo-style training — where GPU clusters train locally and sync intermittently — found that matched random deferral, which picks sync windows at random while matching the same total sync budget, ties or beats every smart, forecast-free scheduling policy tested. That's a problem for a field that typically compares against fixed-period syncs, not random ones. The team introduced Workload-Aware DiLoCo (WA-DiLoCo), a score-based controller that weighs learner progress against fleet load, along with a calibration protocol to pinpoint when a smart policy can actually beat random. They validated it on real vLLM serving traffic replays.
The practical payoff is narrow but statistically real: adding a one-step EWMA burst forecast cuts service-level objective violations from 6.54% to 5.09% in bursty traffic — eight of ten seeds, p=0.021. More broadly, the paper argues that any claim of serving-SLO improvement on shared AI infrastructure should come with a matched-random envelope and real sidecar replay results, not just a fixed-period baseline.
It lands somewhere between methodological rebuke and modest engineering win — researchers with sync-scheduling papers under review may want to check their baselines before anyone else does.