[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"detail-sidebar-cat-0-en-105":3,"doc-seo-273429-105":59,"doc-detail-273429-en":130},{"code":4,"msg":5,"data":6},0,"success",[7,13,18,23,28,33,38,43,48,51,55],{"id":8,"doc_module":4,"doc_module_name":9,"category_name":10,"show_sort_weight":11,"slug":12},1,"Document","Story & Novel",90,"story-novel",{"id":14,"doc_module":4,"doc_module_name":9,"category_name":15,"show_sort_weight":16,"slug":17},2,"Literature",80,"literature",{"id":19,"doc_module":4,"doc_module_name":9,"category_name":20,"show_sort_weight":21,"slug":22},4,"Exam",70,"exam",{"id":24,"doc_module":4,"doc_module_name":9,"category_name":25,"show_sort_weight":26,"slug":27},5,"Comic",60,"comic",{"id":29,"doc_module":4,"doc_module_name":9,"category_name":30,"show_sort_weight":31,"slug":32},6,"Technology",50,"technology",{"id":34,"doc_module":4,"doc_module_name":9,"category_name":35,"show_sort_weight":36,"slug":37},7,"Healthcare",40,"healthcare",{"id":39,"doc_module":4,"doc_module_name":9,"category_name":40,"show_sort_weight":41,"slug":42},8,"Research & Report",30,"research-report",{"id":44,"doc_module":4,"doc_module_name":9,"category_name":45,"show_sort_weight":46,"slug":47},9,"Religion & Spirituality",20,"religion-spirituality",{"id":46,"doc_module":4,"doc_module_name":9,"category_name":49,"show_sort_weight":46,"slug":50},"World Cup","world-cup",{"id":52,"doc_module":4,"doc_module_name":9,"category_name":53,"show_sort_weight":52,"slug":54},10,"Lifestyle","lifestyle",{"id":56,"doc_module":4,"doc_module_name":9,"category_name":57,"show_sort_weight":24,"slug":58},19,"General","general",{"code":4,"msg":60,"data":61},"ok",{"site_id":62,"language":63,"slug":64,"title":65,"keywords":66,"description":67,"schema_data":68,"social_meta":123,"head_meta":125,"extra_data":127,"updated_unix":129},105,"en","training-agents-to-evolve-with-their-harness-taolive-digital-avatar-agent-technical-report","Training Agents to Evolve with Their Harness - TaoLive Digital Avatar Agent Technical Report","","Live e-commerce digital avatar agents must answer product questions, converse with viewers, and execute marketing strategies in real time, making low latency, frequent strategy updates, and reliable responses essential. The report addresses evolvable Harness architectures where skills, hooks, prompts, and tools change without retraining model weights. It proposes Harness-Aware Training (HAT) with Harness-State Augmentation and a three-stage training pipeline (HSA-SFT, OPD, HSA-RL). Evaluations show strong Live-Stream QA and Harness-Variant QA results, avoided IFEval degradation, and real-time latency on a single NVIDIA H20 GPU.",{"@graph":69,"@context":122},[70,84,105],{"@type":71,"itemListElement":72},"BreadcrumbList",[73,77,79,82],{"item":74,"name":75,"@type":76,"position":8},"https://docshare.wps.com","Home","ListItem",{"item":78,"name":9,"@type":76,"position":14},"https://docshare.wps.com/document/",{"item":80,"name":40,"@type":76,"position":81},"https://docshare.wps.com/document/research-report/",3,{"item":83,"name":65,"@type":76,"position":19},"https://docshare.wps.com/document/training-agents-to-evolve-with-their-harness-taolive-digital-avatar-agent-technical-report/273429/",{"url":83,"name":65,"@type":85,"image":86,"author":91,"headline":65,"publisher":94,"fileFormat":97,"inLanguage":63,"description":67,"dateModified":98,"datePublished":99,"encodingFormat":97,"isAccessibleForFree":100,"interactionStatistic":101},"DigitalDocument",{"url":87,"@type":88,"width":89,"height":90},"https://docshare.wps.com/thumbnails/training-agents-to-evolve-with-their-harness-taolive-digital-avatar-agent-technical-report/273429.png","ImageObject",300,407,{"name":92,"@type":93},"Finn","Person",{"url":74,"name":95,"@type":96},"DocShare","Organization","application/pdf","2026-11-01","2026-09-15",true,{"@type":102,"interactionType":103,"userInteractionCount":56},"InteractionCounter",{"@type":104},"ViewAction",{"@type":106,"mainEntity":107},"FAQPage",[108,114,118],{"name":109,"@type":110,"acceptedAnswer":111},"What challenge does the report address for live e-commerce digital avatar agents?","Question",{"text":112,"@type":113},"It targets the need for ultra-low latency while supporting an evolvable Harness that updates skills, hooks, prompts, and tools frequently. The main difficulty is avoiding performance collapse when the Harness changes.","Answer",{"name":115,"@type":110,"acceptedAnswer":116},"How does Harness-Aware Training (HAT) help the model evolve with the Harness?",{"text":117,"@type":113},"HAT trains a compact model to understand the current Harness rather than memorize a fixed configuration. It uses Harness-State Augmentation (HSA) to vary skill identifiers/content, tool schemas, prompt structure, and hook functions during training.",{"name":119,"@type":110,"acceptedAnswer":120},"What are the three stages of HAT and what is the purpose of each stage?",{"text":121,"@type":113},"Stage 1 (HSA-SFT) improves reasoning and tool-calling using high-quality trajectories across diverse environments. Stage 2 (General OPD) uses On-Policy Distillation to restore generalizability. Stage 3 (HSA-RL) applies reinforcement learning in augmented environments to further strengthen robustness to Harness changes and tool usage.","https://schema.org",{"og:url":83,"og:type":124,"og:title":65,"og:site_name":95,"og:description":67},"article",{"robots":126,"canonical":83},"index,follow",{"doc_id":128,"site_id":62},273429,1789471388,{"code":4,"msg":5,"data":131},{"doc_id":128,"user_id":132,"nickname":92,"user_avatar":133,"doc_module":4,"category_id":39,"category_name":40,"doc_title":65,"doc_description":67,"doc_content":134,"file_id":135,"file_url":136,"file_type":137,"file_size":138,"view_count":56,"is_deleted":4,"is_public":8,"is_downloadable":8,"audit_status":8,"page_count":139,"language":140,"language_code":63,"site_id":62,"html_lang":63,"table_of_contents":141,"faqs":142,"seo_title":143,"seo_description":67,"update_tm":129,"read_time":144,"preview_page_count":4},34359740700684,"https://ap-avatar.wpscdn.com/avatar/1f400023980c374ae676?_k=1777273430885731487","arXiv : 2608 . 15763v4 [cs .CL] 1 1 Sep 2026  \nTraining Agents to Evolve with Their Harness: TaoLive Digital Avatar Agent Technical Report  \nTaoLive AIGC LLM Team  \nABSTRACT  \nLive e-commerce digital avatar streamers must answer product questions, interact with viewers, and execute marketing strategies in real time, requiring low latency, frequent strategy updates, and accurate yet effective replies. To meet the need for frequent updates, a common practice is to introduce an evolvable Harness, where Skills, Hooks, prompts, and tools can be modified independently of model weights. Ideally, this allows the system to iterate without retraining the model for every strategy change. However, this introduces a dilemma. While large models possess the generalization required to adapt to dynamic Harnesses zero-shot, their high latency prohibits real-time interaction. Conversely, low-latency compact models, typically trained on a fixed Harness, overfit to specific configurations and fail to evolve with Harness updates without continuous retraining. We propose Harness-Aware Training (HAT) to train a compact model that can evolve with the Harness. The core idea is to expose the model to diverse Harness configurations during training so that it learns to understand the current Harness rather than memorize a fixed version. HAT introduces Harness-State Augmentation (HSA), which applies task-preserving transformations to Skill identifiers, Skill content, tool schemas, prompt structure, and Hook functions. Based on HSA, the training is organized into three stages: (1) HSA-SFT: the compact model learns from high-quality trajectories generated by strong models in diverse environments, directly improving its reasoning and tool-calling abilities. (2) General OPD: the model learns from the base model on general data via On-Policy Distillation to recover the generalizability damaged by SFT. (3) HSA-RL: reinforcement learning is applied in diverse augmented environments to further improve the model’s understanding of changing Harnesses and its tool-calling skills, enhancing overall robustness. Across four complementary evaluation sets, the HAT-trained model reaches an average score of 94.8 on Live-Stream QA (base 80.3, strongest general LLM 93.0) and 94.6 on Harness-Variant QA (base 75.4) . Crucially, while traditional Fixed-Harness SFT causes a significant 7.7-point drop from the base model on IFEval, our method successfully avoids this degradation, achieving a strong score of 83.5 . Furthermore, deployed on a single NVIDIA H20 GPU (with optimization enabled), it achieves a P50 latency of 3.4 s and P95 of 8.1 s, fully satisfying real-time deployment constraints. The system is also deployed in Taobao Live’s production digital-avatar service. In an online A/B test, the Harness treatment recorded a UV-normalized uplift of 0.91% in item-page views relative to the ReAct control.  \n1 Introduction   \nDigital avatars are increasingly deployed across diverse interactive applications, spanning conversational assistants, customer service, education, and marketing [Cui and Liu, 2023, Allal-Chérif et al., 2024, Yang et al., 2025] . Among these, live-streaming e-commerce represents a uniquely demanding scenario where AI-powered streamers must interact with viewers, answer product queries, and execute marketing strategies under strict real-time latency constraints [Liu et al., 2025, Sun et al., 2025, Yu et al., 2026] . Since our live-streaming digital avatar’s replies are broadcast publicly, the underlying agent must simultaneously satisfy three stringent requirements: (1) ultra-low latency, essential for maintaining viewer engagement. (2) high adaptability to frequent shifts in campaign rules, compliance and merchant preferences. (3) accuracy and effectiveness, ensuring factually grounded responses that successfully address viewer intent without hallucination.  \nTo achieve this agility without lengthy model retraining cycles, we adopt an evolvable Harness architec","cbCaihQc1nmiunDd","https://ap.wps.com/l/cbCaihQc1nmiunDd","pdf",3265089,28,"English","# Introduction\n## Problem: real-time requirements and evolvable Harnesses\n## Proposed solution: Harness-Aware Training (HAT)\n## Training stages: HSA-SFT, OPD, HSA-RL\n## Evaluation results and deployment latency","[{\"question\":\"What challenge does the report address for live e-commerce digital avatar agents?\",\"answer\":\"It targets the need for ultra-low latency while supporting an evolvable Harness that updates skills, hooks, prompts, and tools frequently. The main difficulty is avoiding performance collapse when the Harness changes.\"},{\"question\":\"How does Harness-Aware Training (HAT) help the model evolve with the Harness?\",\"answer\":\"HAT trains a compact model to understand the current Harness rather than memorize a fixed configuration. It uses Harness-State Augmentation (HSA) to vary skill identifiers/content, tool schemas, prompt structure, and hook functions during training.\"},{\"question\":\"What are the three stages of HAT and what is the purpose of each stage?\",\"answer\":\"Stage 1 (HSA-SFT) improves reasoning and tool-calling using high-quality trajectories across diverse environments. Stage 2 (General OPD) uses On-Policy Distillation to restore generalizability. Stage 3 (HSA-RL) applies reinforcement learning in augmented environments to further strengthen robustness to Harness changes and tool usage.\"}]","Training Agents to Evolve with Their Harness - TaoLive Digital Avatar Agent Technical Report | PDF",71]