A technical report from Alibaba’s Taobao Live team describes Harness-Aware Training (HAT), a method for training compact models to adapt to frequently-updated agent harnesses used by AI digital-avatar streamers. HAT combines supervised fine-tuning on harness-state augmented data, on-policy distillation, and reinforcement learning to keep small models robust as skills, prompts, and tools change independently of model weights. Deployed in Taobao Live’s production digital-avatar service on a single NVIDIA H20 GPU, the system achieves 3.4s/8.1s P50/P95 latency and produced positive online A/B test results for GMV and item-page views.
