Training compact agents to read their runtime config beats both retraining and zero-shot

The core problem here is deceptively simple: a live-streaming AI avatar needs to respond in real time, but the business rules it operates under change constantly. Retraining a model every time a merchant updates their sales strategy is untenable. Using a large model that can adapt zero-shot is equally untenable—DeepSeek-V4-flash clocks a median end-to-end latency above 11 seconds, which kills viewer engagement. So the system is stuck between a fast model that overfits and a flexible model that's too slow.

The answer proposed here is Harness-Aware Training (HAT), built around a modular runtime called a Harness—four independently versioned components: Skills (reply rules and strategies), Hooks (validation and retry logic), a System Prompt Pipeline, and a Tool Registry. The insight is that if you train a compact model on trajectories collected under only one Harness configuration, it memorizes surface forms—specific skill names, tool schemas, prompt templates—rather than learning to read whatever configuration it's currently given. When the Harness evolves, the model breaks.

HAT fixes this by making Harness variation part of the training distribution itself. Harness-State Augmentation (HSA) applies task-preserving perturbations across five dimensions: skill identifiers, skill content, tool definitions, system prompt structure, and Hook behavior. Training then proceeds in three stages: supervised fine-tuning on teacher-generated trajectories across augmented Harness variants, an on-policy distillation step against the base model on general data to recover instruction-following ability lost during domain SFT, and finally reinforcement learning inside a production-informed simulator running under augmented Harness states.

The numbers are convincing. Fixed-Harness SFT causes a 7.7-point drop on IFEval; HAT avoids this entirely, scoring 83.5. On Live-Stream QA, HAT reaches 94.8 average, beating both the untuned base model at 80.3 and the strongest zero-shot model tested at 93.0. The gap between performance on the original Harness and a variant Harness is only 0.2 points for HAT, versus 4.9 points for the base model. Running on a single H20 GPU with speculative decoding enabled, P50 wall-clock latency sits at 3.4 seconds and P95 at 8.1 seconds.

The production A/B test on Taobao Live—over 1.9 million unique visitors across both arms—shows UV-normalized uplifts of 4.33% in confirmed-receipt GMV and 0.91% in item-page views against the ReAct control. The paper is careful to note that this measures the full system contrast, not the model in isolation.

The ablation work is particularly useful. RL without HSA actually degrades Prompt Robustness below the base model's score, demonstrating that reinforcement learning in a fixed environment amplifies overfitting rather than curing it. The on-policy distillation stage and HSA are complementary: one recovers general capabilities, the other prevents Harness-specific memorization from forming in the first place.