Engineering and algorithmic interventions for multimodal post-training at Microsoft scale

Microsoft said post-training for its large-scale multimodal Copilot agents became unstable in real production settings.

Standard policy-gradient training started degrading long-horizon execution, robustness, and task completion even when reward metrics looked healthy.

The team separated verifiable checks from preference judgments and added five stabilization interventions to keep advantage estimates informative.

The findings show that agent training methods that work in labs can fail at production scale under latency and safety constraints.