Industrial adoption of post-training methods under live-system constraints

Establish how published post-training methods can be safely adopted on live, deployed systems while satisfying fixed-budget, regression-safety, and integration constraints, thereby closing the gap between post-training research and industrial practice.

Background

The paper characterizes industrial post-training as brownfield maintenance: teams begin with a deployed checkpoint, operate under fixed compute and mixture budgets, and must improve a targeted capability without degrading existing capabilities. In this setting, new data must displace or reweight existing data, synthetic-data generation is limited by the yield of usable supervision, and local mixture changes can produce non-local behavioral regressions.

Although the case study demonstrates a successful failure-driven-synthesis patch for competitive-programming code generation, the paper presents the broader unresolved issue as the gap between published research techniques and the practical requirements of safely integrating those techniques into a live system. The authors frame this as an engineering and methodological problem involving mixture accounting, yield measurement, probabilistic regression testing, efficient experimental design, and conservative patch integration.