Training Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical Report
TaoLive AIGC LLM Team, Yuhan Sun, Wenhao Lin, Yongdong Luo, Yibo Hu, Meiguang Jin, Junfeng Ma, Weihang Pan, Jiaxin Zhao, Zulong Chen
cs.CL
2026-08-16
HAT trains Qwen3.6-35B-A3B to follow changing Harnesses, scoring 94.8 Live-Stream QA vs 93.0 for GLM-5.2, with +4.33% GMV per UV on Taobao Live.
A Taobao Live digital-avatar host has to answer product questions, take comments, and run marketing tactics in real time. The business rules change weekly: new Skills, new tool schemas, new prompts, new Hooks. Large models can follow those Harness edits zero-shot, but they are too slow for live streaming. Compact models meet the latency budget and then overfit the current Skill names, tool names, and prompt templates, so the next Harness update breaks them.
Fixed-Harness SFT writes that failure into the weights. On Qwen3.6-35B-A3B, IFEval prompt-level accuracy falls from 81.5 to 73.8, a 7.7-point drop. Operators can rewrite a Skill; the model still routes by the old string.
Harness-Aware Training (HAT) treats Harness variation as part of the training distribution. The Harness state is ℎ=(S,T,P,K): the Skill set, the tool registry, the assembled system prompt, and lifecycle Hooks. Harness-State Augmentation (HSA) keeps task semantics and changes the surface form.
Training has three stages. HSA-SFT: a strong teacher rolls out on original and augmented Harnesses; trajectories are filtered by Accuracy and Effectiveness, then used for supervised fine-tuning. Most Hook-retry traces keep only the final correct behavior, with a small fraction of full retries retained for recovery. General OPD: the pre-SFT base model teaches on Tulu3 via on-policy distillation to restore instruction following. HSA-RL: GRPO/GDPO in a live-stream simulator under HSA, with rewards for Accuracy, Effectiveness, Tool Rationality, and Skill Selection, plus a CoT-length penalty in a 100–200 token linear band.
The backbone is Qwen3.6-35B-A3B, with 10K real live-stream SFT examples and 4K RL interaction tasks.
HAT lifts domain quality without the usual instruction-following collapse.
| Setup | Live-Stream QA | Harness-Variant QA | IFE-P |
| Base | 80.3 | 75.4 | 81.5 |
| Fixed-Harness SFT | 89.5 | 88.2 | 73.8 |
| HAT | 94.8 | 94.6 | 83.5 |
| Strongest general LLM (GLM-5.2) | 93.0 | 93.5 | 89.6 |
T1 and T2 differ by 0.2 points, versus 4.9 for the base. The same held-out Harness edits cut detected errors by 18.1% for Fixed-SFT and 51.7% for HAT. On one H20 with MTP, concurrency 1 yields P50 3.407 s and P95 8.114 s, with all 100 replayed requests under 15 s. A production A/B on Taobao Live (1.58M control UVs, 395K treatment) shows +4.334% UV-normalized confirmed-receipt GMV and +0.911% item-page views. In a 100-request human blind test, Harness wins 35, ties 64, ReAct wins 1.
This is a shippable systems report, not another compact-model distillation note. It puts a moving runtime into the training objective so a latency-feasible 35B-A3B MoE can beat several API-scale models on the business sets without wrecking IFEval. For any agent that hot-swaps Skills, tools, and prompts, HSA is a more targeted fix than repeated full SFT.
There is no dedicated Limitations section. The online A/B compares a full production stack against ReAct, so it does not isolate the policy, the Harness, or routing. The platform labels both metrics “significantly positive” without p-values or intervals, and the snapshot records one significant day each. Offline scores come from an Agent-as-a-Judge calibrated on 482 human labels; the blind test agrees in direction, but it is still a model judging a model. Training-dynamics plots are single-run. MTP helps only at low concurrency; at C=4–8, P95 stops improving.