Alibaba trains compact avatar agent to 94.8 QA on Taobao Live

Alibaba's Taobao Live team has published a technical report on the digital-avatar streamer running in production on its shopping platform, describing a training approach that lets a compact model keep pace with the scaffolding shifting around it. They report 94.8 on their Live-Stream QA benchmark against an 80.3 base and 93.0 for what they call the "strongest general LLM," with a P50 latency of 3.4 seconds on a single NVIDIA H20 GPU.

The technique they name is Harness-Aware Training. The paper frames the problem in terms most agent teams will recognize: "Evolvable Harnesses, whose Skills, Hooks, prompts, and tools can be updated independently of model weights, enable rapid iteration but expose a trade-off: large models adapt zero-shot yet are too slow, whereas compact models meet latency targets but overfit to fixed Harness configurations." Their fix is Harness-State Augmentation, which applies "task-preserving transformations to Skill identifiers and content, tool schemas, prompt structures, and Hook functions," run through a three-stage pipeline of SFT, on-policy distillation, and reinforcement learning.

On a separate Harness-Variant QA test built for this setting, they report 94.6 against a 75.4 base, and they claim the approach avoids a 7.7-point IFEval regression that fixed-harness SFT introduces, landing at 83.5. The report says deployment in Taobao Live's digital-avatar service "yields positive online A/B test results for GMV and item-page views," without publishing the lift figures.