TaoLive Digital Avatar Agent Technical Report: Training Agents to Evolve with Their Harness

# TaoLive Digital Avatar Agent Technical Report:Training Agents to Evolve with Their Harness
Source: [https://arxiv.org/html/2608.15763](https://arxiv.org/html/2608.15763)
∗Equal contribution

###### Abstract

AI\-powered digital avatar streamers in live e\-commerce 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 enable frequent updates, we build an evolvable Harness whose Skills, Hooks, prompts, and tools can be modified independently of model weights, allowing the system to iterate without retraining\. However, not all models can benefit from Harness Evolution\. Models with weak generalization—such as those trained via SFT on a single fixed Harness—overfit to the fixed configuration and cannot evolve with the Harness, while large models with strong generalization have latencies too high for real\-time interaction\. 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, and organizes training into three stages: SFT with HSA \(HSA\-SFT\), General Domain On\-Policy Distillation \(General OPD\), and Agentic RL with HSA \(HSA\-RL\)\. 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, it maintains 83\.5 on IFEval, avoiding the significant 7\.7\-point drop typically observed after Fixed\-Harness SFT\. On a single NVIDIA H20 GPU \(with optimization enabled\), it achieves a P50 latency of 3\.407 s and P95 of 8\.114 s\. These results demonstrate that HAT produces a compact agent that evolves with ongoing Harness changes under strict latency constraints\.

## 1 Introduction

Digital avatars, also referred to as virtual humans are increasingly deployed across diverse interactive applications including conversational assistants, customer service, education, and marketing\([Cui and Liu 2023](https://arxiv.org/html/2608.15763#bib.bib9);[Allal\-Chérif et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib3);[Yang et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib43)\)\. Among these, live\-streaming e\-commerce represents a uniquely demanding scenario: AI\-powered digital streamers must interact with viewers in real\-time, answer product questions, and execute marketing strategies under strict latency constraints\([Liu et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib22);[Sun et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib37)\)\. In our application, the digital avatar’s replies are spoken publicly to the entire broadcast room, requiring the underlying agent to simultaneously satisfy three conflicting requirements\. First, ultra\-low latency is essential, as long wait times severely degrade the live viewing experience\. Second, the agent demands high adaptability; campaign rules, compliance requirements, and reply styles change frequently driven by the shifting preferences of operations teams and merchants\. Third, replies must be both accurate and effective—maintaining factual groundedness without making unsupported claims \(hallucinations\), while successfully addressing viewer intent\.

While a standard ReAct\-style loop\([Yao et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib44)\)enables reasoning and tool usage, it does not facilitate agile strategy updates\. In initial deployments, addressing new failure cases or modifying merchant\-specific strategies often required altering central prompts or control code, followed by a full service deployment or even model retraining\. This slow iteration cycle cannot keep pace with the high\-frequency demands of live commerce\. To address this, we propose an evolvable Harness architecture composed of four independently updatable modules:Skills\(reply rules and strategies\),Hooks\(intermediate step validation and retry logic\), aSystem Prompt Pipeline\(dynamic instruction assembly based on context\), and aTool Registry\(commerce tool definitions\)\. This architecture allows the execution environment to evolve while policy model weights remain fixed—a practice we term Harness Evolution, which is closely related to modular prompt and runtime optimization\([Khattab et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib15);[Hebbar et al\. 2026](https://arxiv.org/html/2608.15763#bib.bib13);[Zhang et al\. 2026](https://arxiv.org/html/2608.15763#bib.bib51);[Lin et al\. 2026](https://arxiv.org/html/2608.15763#bib.bib20);[Qu et al\. 2026](https://arxiv.org/html/2608.15763#bib.bib31);[Ning et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib26)\)\. Our system adopts this approach through a developer\-reviewed diagnose–edit–evaluate loop applied to these modules, enabling business behavior to be adjusted within hours rather than requiring a lengthy retraining cycle\. \(Section[2\.3](https://arxiv.org/html/2608.15763#S2.SS3)\)\.

Harness Evolution, however, changes the problem that the policy model must solve\. Since every promoted Harness version replaces the configuration the model was initially deployed against, the model no longer acts inside a fixed execution environment but rather inside a continuous sequence of evolving environments: Skills are rewritten, Hooks are added, prompts are reassembled, and tool schemas are renamed, all while the weights stay unchanged\. Robustness to Harness changes is therefore a property the policy model must possess rather than a convenience we obtain for free; ironically, the more effective the evolution loop becomes, the more frequently the model’s environment moves beneath it\.

Which model can inhabit such a moving environment? Strong zero\-shot models are the natural candidates, as their general instruction\-following lets them absorb a rewritten Skill or a renamed tool with no adaptation at all—but their latency rules them out, with median end\-to\-end latencies exceeding 11 seconds for DeepSeek\-V4 models in our tests, far beyond the real\-time budget\. The latency requirement thus forces a compact policy model \(e\.g\., Qwen3\.6\-35B\-A3B\), which in turn must be domain\-trained, because its zero\-shot accuracy does not meet industrial standards\. This is where the two mechanisms collide: Fixed\-Harness SFT on the single Harness configuration available at training time yields surface\-form overfitting—the model memorizes specific Skill names, tool names and prompt templates instead of reading the instructions actually placed in front of it—and that configuration is precisely what Harness Evolution is designed to replace\. A model trained this way becomes stale exactly when the system evolves successfully\. Our experiments make this cost concrete: Fixed\-Harness SFT improved domain\-specific metrics but degraded general instruction\-following ability \(measured by IFEval\([Zhou et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib54)\)\) by 7\.7 points, the very capability needed to follow an unfamiliar Harness\. Nor do existing alternatives resolve the tension: test\-time adaptation\([Liang et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib19);[Chen et al\. 2026a](https://arxiv.org/html/2608.15763#bib.bib4)\)violates strict latency budgets, while continual learning\([Wang et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib41)\)re\-pays a retraining cost for every Harness edit, which is unsustainable at our update frequency\.

We resolve this tension with Harness\-Aware Training \(HAT\), which makes the Harness state an explicit part of the training distribution rather than a fixed background assumption\. Instead of fitting the policy to trajectories collected under the single configurationhth\_\{t\}that happens to be in production, we train over a distribution of Harness states, so that the model learns to condition on whatever Harness it is currently given and can evolve together with it\. Concretely, we introduce Harness\-State Augmentation \(HSA\) to apply task\-preserving perturbations to Skill content, tool schemas, prompt structure, and interaction constraints, and instantiate it in two training stages\. The complete pipeline comprises \(1\)SFT with HSA \(HSA\-SFT\)using trajectories generated under varied Harness configurations; \(2\)General On\-Policy Distillation \(General OPD\)\([Agarwal et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib2)\)to recover general capabilities degraded during domain\-specific SFT\([Lu and Lab 2025](https://arxiv.org/html/2608.15763#bib.bib24)\); and \(3\)Agentic RL with HSA \(HSA\-RL\)within a production\-informed live\-room simulator, allowing the model to learn from the consequences of its actions under shifting environmental conditions\([Zhou et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib55);[Jin et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib14)\)\.

Extensive evaluations demonstrate the effectiveness of our approach\. On our real\-world Live\-Stream QA benchmark, the selected 35B checkpoint achieves a 94\.8 average score, surpassing both the untuned base model \(80\.3\) and strong zero\-shot models\. Crucially, it maintains robustness to evaluated Harness changes and avoids the significant IFEval score loss observed after Fixed\-Harness SFT\.

Our Harness\-based judge is calibrated against human annotations to ensure reliability\. In controlled deployment tests on a single NVIDIA\-H20 GPU with MTP\([DeepSeek\-AI 2024](https://arxiv.org/html/2608.15763#bib.bib10);[Li et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib18);[Chen et al\. 2026b](https://arxiv.org/html/2608.15763#bib.bib6);[Cheng et al\. 2026](https://arxiv.org/html/2608.15763#bib.bib8)\), the complete agent meets strict latency requirements, achieving a P50 wall\-clock latency of 3\.407s and a P95 of 8\.114s\.

Our contributions are:

- ▶\\blacktrianglerightAn evolvable Harness, and the training problem it creates\.We decouple Skills, prompts, Hooks, and tools from the policy model, enabling agile updates to strategy and constraints without retraining; Harness Evolution alone significantly improves success rates on our development set\. We then characterize its consequence: a successful evolution loop turns the policy model’s execution environment into a moving target that fixed\-Harness training cannot accommodate \(Section[2](https://arxiv.org/html/2608.15763#S2)\)\.
- ▶\\blacktrianglerightHarness\-Aware Training: putting the Harness into the training distribution\.We introduce HSA and instantiate it in a three\-stage HAT pipeline comprising HSA\-SFT, General OPD, and HSA\-RL\. This design lets a latency\-feasible compact model track Harness changes instead of memorizing one configuration while retaining general capabilities \(Sections[3](https://arxiv.org/html/2608.15763#S3)and[5\.4](https://arxiv.org/html/2608.15763#S5.SS4)\)\.
- ▶\\blacktrianglerightComprehensive benchmarks and reliable evaluation procedures\.We construct four complementary evaluation sets covering real\-world interactions, Harness variations, and synthetic robustness scenarios, totaling over 4,500 cases\. We validate our Harness\-based Agent\-as\-a\-Judge against human labels to ensure reliable, scalable evaluation \(Section[4](https://arxiv.org/html/2608.15763#S4)\)\.
- ▶\\blacktrianglerightPractical insights for training agents that can evolve with their Harness\.Our extensive experiments reveal critical findings, such as the larger observed impact of HSA during SFT than during RL, and the role of General OPD in recovering capabilities lost during domain adaptation \(Section[5\.3](https://arxiv.org/html/2608.15763#S5.SS3)\)\.

## 2 Digital\-Avatar Harness Agent Architecture

The Harness Agent separates a slowly updated*policy model*from a rapidly evolving*Harness state*\. This separation lets operators change business behavior without retraining, but it also creates the central learning problem of this paper: the policy must remain effective as the environment around it evolves\. We first describe the application and runtime, then formalize the non\-training process as*Harness Evolution*\.

### 2\.1 Live\-Streaming Digital\-Avatar Scenario

We study a digital\-avatar host that answers product questions, handles open\-ended conversation, and invokes external tools during live commerce\. Public spoken responses are synthesized through TTS and avatar driving for the whole room\. Each request combines the viewer message with anchor speech, product data, link IDs, and living\-room status\. The application demands low latency while business rules evolve weekly\.

Figure 1:Application flow and the two independent update paths\.*Harness Evolution*edits agent runtime modules while keeping model weights fixed;*Harness\-Aware Training \(HAT\)*provides a well\-performing, low\-latency, and robust model\. Both paths meet inside the same production runtime\.
### 2\.2 The Harness Agent Runtime

The Harness Agent runtime comprises four mechanisms\.*Skills*are dynamically loaded behavioral modules for Q&A strategy, tool choosing, and reply style\. The*System Prompt Pipeline*assembles instructions from product specifications, retrieved FAQ, living\-room status, global boundary and active skill descriptions\.*Hooks*validate inputs, parameters, and output formats, triggering correction or retry when needed\.*Tool orchestration*connects the agent to external product, inventory, and commerce services\. Because these modules are versioned independently of model weights, a new skill or tool can be deployed as a harness change rather than a training cycle\. And our harness agent architecture is shown in Fig\.[3](https://arxiv.org/html/2608.15763#S2.F3)\.

### 2\.3 Harness Evolution without Weight Updates

We useHarness Evolutionto denote a controlled process that improves the agent by changing skills, prompts, hooks, or tools while holding the model fixed\. It is human\-in\-loop rather than fully autonomous: AI performs failure clustering, diagnosis, and proposed edits; a developer confirms the plan, runs evaluation, and decides whether to promote, revise, or stop\. The complete operational protocol appears in Appendix[B](https://arxiv.org/html/2608.15763#A2)\.

Harness Evolution loopDiagnose→\\rightarrowConfirm→\\rightarrowEdit Harness→\\rightarrowEvaluate→\\rightarrowRegression Check

Figure[2](https://arxiv.org/html/2608.15763#S2.F2)illustrates both the value and the limit of the harness evolution process\. On the 482\-case human\-labeled calibration set annotated by three annotators, moving from ReAct to the modular harness raises Effectiveness from 84\.58 to 87\.16\. Evolution 1 sharply improves Accuracy from 82\.40 to 92\.13 but reduces Effectiveness to 84\.16\. Evolution 2 reaches 92\.55 Accuracy and 92\.75 Effectiveness\. Additional long\-tail rules in Evolutions 3–4 regress one or both dimensions, so Evolution 2 is selected as the engineering early\-stop checkpoint\. Detailed module changes, category diagnostics, stopping logic, and evidence boundaries are reported in Appendix[B](https://arxiv.org/html/2608.15763#A2)\.

Figure 2:Harness Evolution with a fixed DeepSeek\-V4\-flash on dev\-set \(n=482n=482\)\. The process is non\-monotonic: Evolution 2 is the selected development checkpoint, while later long\-tail edits introduce regressions\.Why evolution changes the training problemEvery promoted harness version creates a new stateht\+1h\_\{t\+1\}while the model remains fixed\. A model that memorizeshth\_\{t\}therefore becomes stale precisely when the system evolves successfully\.

The evaluation Judge is also implemented as an editable harness agent as detailed in Appendix[D](https://arxiv.org/html/2608.15763#A4)\.

### 2\.4 Task and Runtime Details

This section expands the task interface and runtime inventory summarized in Section[2](https://arxiv.org/html/2608.15763#S2)\. Evaluation roles and scoring are specified separately in Appendix[E](https://arxiv.org/html/2608.15763#A5)\.

#### Task Interface

The agent operates in live\-streaming e\-commerce, where a digital\-avatar host responds to viewer comments under the production latency bound\. Each request contains the viewer’s comment, the anchor’s nearby utterances, live\-room and product context, applicable FAQ entries, and recent dialogue history\. The output contains a natural\-language reply—either public speech sent to TTS and avatar driving or a private text response—plus the structured tool\-call trajectory used to produce it\.

In single\-comment mode, the runtime handles one viewer message independently; in multi\-comment mode, it aggregates related messages before responding\. An available internal traffic summary groups the main viewer intents as shown in Table[1](https://arxiv.org/html/2608.15763#S2.T1.fig1)\. The values are rounded, and the archived summary does not preserve the reporting window; they therefore describe workload composition rather than a population estimate\.

Table 1:Approximate intent mix in the available internal traffic summary\.
#### Runtime Interfaces

The Harness exposes built\-in interfaces for retrieval, commerce operations, and flow control\. Additional marketing interfaces are discovered dynamically from an external service through the Model Context Protocol \(MCP\)\. Table[12](https://arxiv.org/html/2608.15763#A1.T12)groups the release inventory by responsibility\. The inventory is bound to manifestHA\-ARL\-20260614, together with the policy checkpoint and Harness version hash\.

#### Skill Routing

Each Skill is a versioned Markdown module with YAML metadata and behavioral directives\. The runtime requires at least one Reply Skill before a substantive final response; Strategy Skills are optional and can be composed when an interaction requires a specialized tool chain\. Table[13](https://arxiv.org/html/2608.15763#A1.T13)records the release inventory\. Module names and versions are reconciled against the same manifest before promotion\.

#### Execution and Trajectory Record

Figure[3](https://arxiv.org/html/2608.15763#S2.F3)summarizes the implementation\-level request path in a deliberately compact form\. It separates the editable Harness modules from the repeated policy–tool loop and distinguishes resilience for an individual model call from recovery of the full trajectory\. Product\-specific tool names and Hook implementations are omitted because they vary across Harness versions\.

Figure 3:Simplified Harness Agent runtime control flow\. Prompt rules and dynamically loaded Skills assemble the model context; lifecycle Hooks validate inference, tool use, and stopping\. A failed stop gate starts another round, while a passed gate produces the final response and structured trajectory\. Per\-call fallback and trajectory\-level recovery protect different failure scopes\.A request traverses the following runtime stages:

1. 1\.Load and normalize context\.Validate the request and assemble product, FAQ, room, and dialogue state\.
2. 2\.Assemble the prompt\.Combine the base instruction, active product and room configuration, relevant retrieved context, and any loaded Skills\.
3. 3\.Run the agent loop\.Under a configurable round limit, the policy can load Skills, issue one or more tool calls, incorporate their results, and propose a final reply\.
4. 4\.Apply Hook checks\.Lifecycle checks cover input validity, tool parameters, Skill\-loading requirements, response format, known factuality risks, and retry control\. The complete deployed Hook manifest remains version\-bound rather than hard\-coded in this description\.
5. 5\.Extract and dispatch the reply\.Remove internal reasoning fields, retain the structured trajectory, and route the public or private response to its downstream channel\.

Example trajectory recordinput\.danmaku = ‘‘How much is product \#3?’’ context\.current\_link\_id = 1 call\[1\] = get\_product\_info\_by\_link\_id\(link\_id=3\) result\[1\] = \{name: ‘‘Sunscreen SPF50’’, price: 89\} call\[2\] = load\_skill\(skill=‘‘item\_qa’’\) reply = ‘‘Link \#3 is Sunscreen SPF50, currently priced at 89 yuan\.’’

## 3 Harness\-Aware Training \(HAT\)

### 3\.1 Problem Formulation

We model a single agent interaction asπθ​\(a∣x,h\)\\pi\_\{\\theta\}\(a\\mid x,h\), wherexxdenotes the user input together with its business context \(e\.g\., live\-streaming room state, current product listing\), andh∈ℋh\\in\\mathcal\{H\}is the*Harness state*—the full configuration that governs agent behavior at inference time\. Concretely,

h=\(𝒮,𝒯,𝒫,𝒦\),h=\(\\mathcal\{S\},\\;\\mathcal\{T\},\\;\\mathcal\{P\},\\;\\mathcal\{K\}\),\(1\)where𝒮\\mathcal\{S\}is the set of active Skills \(pluggable policy modules with reply rules and examples\),𝒯\\mathcal\{T\}is the tool registry \(schemas, parameter definitions, and descriptions\),𝒫\\mathcal\{P\}is the system prompt assembled by the dynamic pipeline, and𝒦\\mathcal\{K\}is the set of Hooks \(checkpoint logic for validation, retry, and format enforcement\)\.

In production,hhevolves continuously: operations teams add Skills weekly, tool interfaces are updated as upstream services change, and prompts are revised in response to emerging bad cases\. A model trained at timettobserveshth\_\{t\}during training but facesht\+1,ht\+2,…h\_\{t\+1\},h\_\{t\+2\},\\ldotsafter deployment\.

Our training objective is twofold: \(1\) improve policy effectiveness when deployment draws a new harnessh′∼ℰdeployh^\{\\prime\}\\sim\\mathcal\{E\}\_\{\\textrm\{deploy\}\}from a specified change envelope that is broader than any single training snapshot; and \(2\) retain general\-purpose capabilities such as instruction following after domain\-specific fine\-tuning\.

### 3\.2 Harness\-State Augmentation \(HSA\)

Our working hypothesis is*surface\-form overfitting*: when training data contains a single, fixed harness, the objective permits shortcuts such as matching skill names by string similarity, selecting tools from memorized schema text, or following instructions only in a familiar template\. The core hypothesis is that changing the surface form of the harness makes it harder for the model to rely on fixed names or templates as shortcuts\. If a skill is renamed, its description rewritten, and distractors injected, functional information about the skill becomes more useful for routing than memorizing a single identifier\.

We perturb the harness along five dimensions, each targeting a distinct aspect of surface\-form dependence:

1. 1\.Skill identifierWe expand the skill set with synthetic plausible skills and noise skills, randomly mask subsets of existing skills, rename skills, and rewrite skill descriptions by a strong language model\. These changes reduce the usefulness of a fixed skill identifier\.
2. 2\.Skill content\.Within each skill, we paraphrase individual rules, randomly mask a subset, and reorder them\. This variation is intended to reduce reliance on rule position and exact wording\.
3. 3\.Tool definitions\.We rename tools, rewrite descriptions\. This transformation makes name matching less reliable while preserving the tool’s function\.
4. 4\.System prompt\.We reorder top\-level instruction blocks and items within blocks, and perturb numeric constraints \(e\.g\., response length limits, maximum interaction rounds\) within operationally valid ranges\.
5. 5\.Hook\.We design controlled variants of retry behavior, message structure, and hook\-triggered modifications to simulate the changes that the model may encounter after hooks are updated in production\.

Figure 4:Overview of HAT\. \(a\) HSA constructs diversified Harness states; HSA\-SFT followed by General OPD produces the shared initialization\. \(b\) Controlled comparison of Agentic RL and Agentic OPD under the original Harness and HSA environments\. Every variant starts from the HSA\-SFT \+ General OPD result, and each OPD variant receives its paired same\-environment RL teacher\. The selected Agentic RL policy under HSA is HSA\-RL\.
### 3\.3 Simulated On\-Policy Environment

HSA addresses environment coverage but operates on*offline*demonstration data: the model learns from teacher\-generated trajectories that always follow the optimal path\. In real deployment, however, the agent must handle situations that never appear in demonstrations—a tool call returns an unexpected error, a Hook intercepts and forces a retry, or the agent selects a wrong Skill and must recover mid\-conversation\. Offline data provide no learning signal from the model’s own failed trajectories and recovery attempts, because they contain only teacher trajectories\.

Moreover, the Harness Agent operates as a multi\-step system: within a single user request, the model may perform Skill selection, tool invocation, result interpretation, and response generation across multiple reasoning rounds, with Hooks potentially triggering retries at each stage\. Learning to navigate this dynamic interaction loop requires the model to actually*experience*it during training—observing the consequences of its own actions rather than imitating a teacher’s\.

##### Environment Design

We construct a production\-informed live\-streaming simulator that implements the control\-flow elements needed for training: skills, tool executors, hooks, and bounded multi\-round interaction\. The simulator consists of four components:

\(1\) Streaming input simulation\.The environment generates realistic user inputs by sampling from a pool of live\-streaming scenarios: product inquiries, casual chat, purchasing intent, after\-sales questions, and mixed\-intent bullet comments\. Each input is contextualized with a simulated room state including current product listings, historical conversation turns, and viewer action signals \(e\.g\., adding to cart, joining the room\)\.

\(2\) Harness Agent scheduling\.The simulator implements the full Harness Agent control flow\. Given the model’s output at each reasoning step, the environment executes Skill routing logic \(determining which Skills are loaded based on the model’s selection\), triggers Hooks at the appropriate checkpoints \(input validation, parameter checking, response format verification, Skill\-loading enforcement\), and manages the multi\-round interaction loop with a configurable maximum round limit\.

\(3\) Tool execution simulation\.When the model invokes a tool, the simulator executes the call against sandboxed service replicas that return product details, pricing information, and inventory status from production catalog snapshots\. The simulator also injects controlled failures—timeout errors, malformed responses, and permission denials—at calibrated rates to expose the model to error\-recovery scenarios\.

\(4\) Augmented Harness configuration\.The simulated environment operates under augmented Harness states produced by the method described in §[3\.2](https://arxiv.org/html/2608.15763#S3.SS2)\. This aligns the offline and on\-policy stages to the same declared change envelope\.

##### Agentic RL Training

Within the simulated environment, we optimize the policy based on Group Relative Policy Optimization \(GRPO\)\([Shao et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib34)\)\. For each prompt,GGcomplete trajectory rollouts \(group size\) are sampled and split into tool segments and reply segments\. The reward signal is computed from four dimensions:

- ▶\\blacktriangleright*Accuracy*: whether the final response contains factually correct information, judged during RL by a DeepSeek\-V4\-flash Reward Judge;
- ▶\\blacktriangleright*Effectiveness*: whether the response adequately addresses the user’s intent with appropriate detail and actionability;
- ▶\\blacktriangleright*Tool Rationality*: whether the tool invocation chain is logically sound and efficiently structured \(no redundant calls, correct parameter usage\);
- ▶\\blacktriangleright*Skill Selection*: whether the model correctly identifies the user’s intent and the current context to load the appropriate Skills\.

As an auxiliary regularizer, we discourage unnecessarily long chain\-of\-thought \(CoT\) spans\. For state groupgg, letLigL\_\{i\}^\{g\}be the sum of tokens in the CoT spans extracted from trajectoryii\. Tool\-call and final\-reply states are measured separately, so one state does not consume the other’s length allowance\. We define

rCoT​\(L\)\\displaystyle r\_\{\\text\{CoT\}\}\(L\)=\{0,L≤Llow,L−LlowLhigh−Llow,Llow<L<Lhigh,1,L≥Lhigh,\\displaystyle=\\begin\{cases\}0,&L\\leq L\_\{\\mathrm\{low\}\},\\\\ \\dfrac\{L\-L\_\{\\mathrm\{low\}\}\}\{L\_\{\\mathrm\{high\}\}\-L\_\{\\mathrm\{low\}\}\},&L\_\{\\mathrm\{low\}\}<L<L\_\{\\mathrm\{high\}\},\\\\ 1,&L\\geq L\_\{\\mathrm\{high\}\},\\end\{cases\}\(2\)HereLlowL\_\{\\mathrm\{low\}\}andLhighL\_\{\\mathrm\{high\}\}are tunable no\-penalty and saturation thresholds, respectively, withLlow<LhighL\_\{\\mathrm\{low\}\}<L\_\{\\mathrm\{high\}\}; we set them to 100 and 200 tokens in the reported experiments\. Thus the raw contribution ranges from 0 to−0\.1\-0\.1per state\. This component is intended to reduce avoidable reasoning length\.

The tool and reply segments are treated as separate groups \(g∈𝒢=\{tool,reply\}g\\in\\mathcal\{G\}=\\\{\\texttt\{tool\},\\texttt\{reply\}\\\}\) for advantage computation\. Since each group carries multiple reward dimensions, we adopt*Group reward Decomposed Policy Optimization*\(GDPO\)\([Liu et al\. 2026](https://arxiv.org/html/2608.15763#bib.bib21)\)to estimate advantages: for each task reward dimensionddwithin state groupgg, a group\-normalized advantageA^i\(d\)=\(ri\(d\)−μg\(d\)\)/σg\(d\)\\hat\{A\}^\{\(d\)\}\_\{i\}=\(r^\{\(d\)\}\_\{i\}\-\\mu^\{\(d\)\}\_\{g\}\)/\\sigma^\{\(d\)\}\_\{g\}is computed\. The CoT score is normalized as its own dimension,A^i\(CoT,g\)=\(q⁡\(Lig\)−μg\(CoT\)\)/σg\(CoT\)\\hat\{A\}^\{\(\\mathrm\{CoT\},g\)\}\_\{i\}=\(q\(L\_\{i\}^\{g\}\)\-\\mu\_\{g\}^\{\(\\mathrm\{CoT\}\)\}\)/\\sigma\_\{g\}^\{\(\\mathrm\{CoT\}\)\}, and its negative weight is applied after this per\-dimension normalization:

A^ig=∑d∈𝒟gA^i\(d\)−λ​A^i\(CoT,g\),A^ig,⋆=A^ig−μgσg\.\\hat\{A\}\_\{i\}^\{g\}=\\sum\_\{d\\in\\mathcal\{D\}\_\{g\}\}\\hat\{A\}\_\{i\}^\{\(d\)\}\-\\lambda\\hat\{A\}^\{\(\\mathrm\{CoT\},g\)\}\_\{i\},\\qquad\\hat\{A\}\_\{i\}^\{g,\\star\}=\\frac\{\\hat\{A\}\_\{i\}^\{g\}\-\\mu^\{g\}\}\{\\sigma^\{g\}\}\.\(3\)Hereμ\\muandσ\\sigmadenote the corresponding within\-group means and standard deviations, and all four task dimensions retain unit weight\. For importance sampling, we adopt the sequence\-level scheme of GSPO\([Zheng et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib52)\), giving the clipped policy gradient objective:

ℒ\\displaystyle\\mathcal\{L\}=−1\|𝒢\|​G∑g∑imin\(ρi\(θ\)A^g,⋆i,clip\(ρi\(θ\),1±ε\)A^g,⋆i\),\\displaystyle=\-\\frac\{1\}\{\|\\mathcal\{G\}\|G\}\\sum\_\{g\}\\sum\_\{i\}\\min\\\!\\Bigl\(\\rho\_\{i\}\(\\theta\)\\,\\hat\{A\}^\{g,\\star\}\_\{i\},\\;\\mathrm\{clip\}\(\\rho\_\{i\}\(\\theta\),1\{\\pm\}\\varepsilon\)\\,\\hat\{A\}^\{g,\\star\}\_\{i\}\\Bigr\),\(4\)ρi​\(θ\)\\displaystyle\\rho\_\{i\}\(\\theta\)=\(πθ​\(yi∣x\)πθold​\(yi∣x\)\)1\|yi\|,\\displaystyle=\\left\(\\frac\{\\pi\_\{\\theta\}\(y\_\{i\}\\mid x\)\}\{\\pi\_\{\\theta\_\{\\mathrm\{old\}\}\}\(y\_\{i\}\\mid x\)\}\\right\)^\{\\\!\\frac\{1\}\{\|y\_\{i\}\|\}\},=exp⁡\(1\|yi\|​∑t=1\|yi\|log⁡πθ​\(yi,t∣x,yi,<t\)πθold​\(yi,t∣x,yi,<t\)\)\.\\displaystyle=\\exp\\\!\\left\(\\frac\{1\}\{\|y\_\{i\}\|\}\\sum\_\{t=1\}^\{\|y\_\{i\}\|\}\\log\\frac\{\\pi\_\{\\theta\}\(y\_\{i,t\}\\mid x,y\_\{i,<t\}\)\}\{\\pi\_\{\\theta\_\{\\mathrm\{old\}\}\}\(y\_\{i,t\}\\mid x,y\_\{i,<t\}\)\}\\right\)\.Hereπθ\\pi\_\{\\theta\}andπθold\\pi\_\{\\theta\_\{\\mathrm\{old\}\}\}are the current and reference policies;yiy\_\{i\}is the generated sequence of length\|yi\|\|y\_\{i\}\|;yi,ty\_\{i,t\}is thett\-th token andyi,<ty\_\{i,<t\}its prefix;xxis the input context; andε\\varepsilonis the PPO clip range\([Schulman et al\. 2017](https://arxiv.org/html/2608.15763#bib.bib33)\)\. The sequence\-level ratioρi​\(θ\)\\rho\_\{i\}\(\\theta\)replaces token\-level importance sampling to stabilize training under the sparse expert routing of the MoE architecture\.

When an individual trajectory has zero advantage after GDPO normalization, it provides no gradient signal while diluting the effective batch\. Such zero\-gradient samples are discarded before optimization\([Yu et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib46)\)\. Algorithm[1](https://arxiv.org/html/2608.15763#alg1)in Appendix[C](https://arxiv.org/html/2608.15763#A3)gives the complete procedure\.

### 3\.4 Training Pipeline

The complete training pipeline consists of three stages, each addressing a distinct capability gap:

##### Stage 1: SFT with Harness\-State Augmentation \(HSA\-SFT\)

HSA\-SFT constructs supervised fine\-tuning samples from real live\-room interactions\. Each original sample contains the viewer query, live\-room context, and product context, and is paired with a corresponding Harness which includes the available skills, tool definitions, system prompt, and Hook configuration\. To reduce the model’s dependence on the surface form of a single harness, we first apply HSA to the Harness base state in Figure[2](https://arxiv.org/html/2608.15763#S2.F2)before generating supervised trajectories\. This produces both the original Harness and multiple augmented Harness variants\. HSA changes skill names and descriptions, skill\-rule order, tool descriptions, system\-prompt structure, and the surface form of Hook messages and retry behavior\.

The teacher model then generates candidate trajectories under both the original harness and the augmented harness variants, so the supervised signals cover multiple simulated environments\. These candidate trajectories are filtered according to Accuracy and Effectiveness\. As a result, supervised learning exposes the student to the same task semantics under different harness realizations, encouraging it to use the functional meaning of the current Harness rather than fixed names or template matches\.

Hooks in production may inform the model about formatting errors, parameter errors, or invalid outputs, and ask it to revise\. We want the model to correct itself when receiving a hook reminder, but we do not want it to learn a default strategy of producing an erroneous output first and relying on retries\. Therefore, for most trajectories that contain hook\-triggered retries, we remove the intermediate failed outputs and hook interactions and keep only the final correct behavior\. We retain a small fraction of complete retry trajectories so that the model still learns how to interpret hook reminders and recover after a failure\.

##### Stage 2: General Domain On\-Policy Distillation \(General OPD\)

Motivated by prior practice showing that On\-Policy Distillation can improve general instruction\-following ability\([Lu and Lab 2025](https://arxiv.org/html/2608.15763#bib.bib24)\), we add a General OPD stage after HSA\-SFT\. We use the pre\-HSA\-SFT base model as the teacher and perform on\-policy distillation on the Tulu3\([Lambert et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib17)\)general instruction dataset\. The student samples responses on\-policy for Tulu inputs and is trained by minimizing the KL divergence between the student distribution and the base\-model distribution:

ℒOPD=Ex∼𝒟gen\[DKL\(pθ\(⋅∣x\)∥pbase\(⋅∣x\)\)\]\.\\mathcal\{L\}\_\{\\textrm\{OPD\}\}=E\_\{x\\sim\\mathcal\{D\}\_\{\\textrm\{gen\}\}\}\\big\[\\,D\_\{\\textrm\{KL\}\}\\\!\\left\(p\_\{\\theta\}\(\\cdot\\mid x\)\\,\\\|\\,p\_\{\\textrm\{base\}\}\(\\cdot\\mid x\)\\right\)\\,\\big\]\.\(5\)

##### Stage 3: Agentic RL with Harness\-State Augmentation \(HSA\-RL\)

From the checkpoint trained by HSA\-SFT and General OPD, we further optimize the policy in the simulated on\-policy environment\. This environment uses the production\-informed simulator defined in Section[3\.3](https://arxiv.org/html/2608.15763#S3.SS3), including live\-streaming scenarios, multi\-round interaction, skill loading, tool execution, and Hooks\. In HSA\-RL, HSA is applied before sequence generation: before each rollout, we first decide whether the current sample uses the original Harness or an augmented Harness, and the model then generates a complete interaction trajectory under that Harness\. In this way, part of the rollouts are exposed to the original production configuration, while the rest are exposed to domain\-preserving augmented Harness states, maintaining adaptation to the original environment while expanding coverage of Harness changes\. For optimization, we use the Agentic RL objective described in Section[3\.3](https://arxiv.org/html/2608.15763#S3.SS3), with the complete training procedure provided in Appendix[C](https://arxiv.org/html/2608.15763#A3)\.

## 4 Evaluation Design

This section describes the evaluation sources, the scoring metrics for each source, and the deployment replay protocol used to measure actual latency and serving performance\.

### 4\.1 Evaluation Sets and Sources

We organize the evaluation around the requirements in the live\-streaming application\. For offline evaluation, Live\-Stream QA and Harness\-Variant QA are both built from real live\-room interactions and cover the main industry scenarios in our application domain, with the available internal intent mix summarized in Appendix Table[1](https://arxiv.org/html/2608.15763#S2.T1.fig1)\. Live\-Stream QA tests industrial reply quality, while Harness\-Variant QA tests generalization and robustness to changing Harness contexts\. We also use a separate real live\-room calibration cohort to align the Harness\-based Agent\-as\-a\-Judge with human labels\. To test broader tool and prompt robustness, we construct synthetic Tool Robustness and Prompt Robustness subsets by expanding real\-industrial scenario seeds with LLM generation and consistency validation, as shown in Appendix[E\.3](https://arxiv.org/html/2608.15763#A5.SS3)\. We also evaluate general instruction\-following ability on the public IFEval benchmark\([Zhou et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib54)\)\. To test actual deployment latency and serving performance, we use a fixed complete\-Agent deployment replay\. Table[2](https://arxiv.org/html/2608.15763#S4.T2)summarizes the size, source, and role of each evaluation source\.

Table 2:Evaluation sets and sources\. The sets cover real live\-stream reply quality, in\-family Harness variation, synthetic tool and prompt robustness, general instruction following, and actual deployment latency\.
### 4\.2 Metrics and Scoring

For Live\-Stream QA and Harness\-Variant QA, we evaluate each response with Accuracy and Effectiveness using the Harness\-based Agent\-as\-a\-Judge, following the LLM\-as\-a\-Judge paradigm\([Zheng et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib53)\)\. Appendix[E\.1](https://arxiv.org/html/2608.15763#A5.SS1)\. Accuracy is a binary metric that penalizes factual errors, unsupported product claims, false promises, and capability overreach\. Effectiveness measures whether the response addresses the viewer’s intent and provides useful information, with scores of 0, 0\.5, or 1\. We report AVG as the per\-sample arithmetic mean of Accuracy and Effectiveness, computed before rounding\.

For Tool Robustness and Prompt Robustness, we use task\-specific rubric metrics\. Tool Robustness uses ADR, RC, and TC: ADR measures whether the response addresses the core viewer query, RC measures whether the response uses the required information returned by tools, and TC measures whether the expected tool is called\. Prompt Robustness uses ADR, LLM, and TCO: ADR keeps the same core\-query definition, LLM evaluates whether the response satisfies the active instruction constraints, and TCO checks whether tool calls follow the required order\.

For IFEval, we use its official evaluator and report prompt\-level and instruction\-level accuracy \(abbreviated as IFE\-P and IFE\-I hereafter\)\. IFE\-P measures whether the model satisfies the full prompt\-level instruction set, while IFE\-I measures whether individual instructions are followed\.

For deployment replay, we report wall\-clock latency P50 and P95, TTFT, decoding throughput, execution success, and 15\-second attainment\. Wall\-clock latency is the time from receiving the input query to obtaining the final Agent answer\. TTFT is the time to the first token\. Execution success indicates completion of the Agent request, not answer quality\. The 15\-second attainment denominator includes all measured requests\.

### 4\.3 Deployment Inference Protocol

We use 110 cases, mark 10 as warm\-up, and compute metrics over the remaining 100 requests\. Every request executes the complete Agent with the same Harness commit, real MCP services, at most four agent loops, and a 5\-second tool timeout\. Generation uses temperature 1\.0 and top\-pp0\.95\. Qwen3\.6\-35B\-A3B and our task\-trained checkpoint run in BF16 on one H20 with TP=PP=1\. Their NextN MTP modules supply draft tokens, which the serving engine processes through the same EAGLE\-style speculative generation and verification path with three speculative steps and four draft tokens; Appendix[F\.1](https://arxiv.org/html/2608.15763#A6.SS1)gives the complete engine configuration\.

## 5 Experiments

### 5\.1 Experimental Setup

The training experiments are conducted on Qwen3\.6\-35B\-A3B\. The training data are collected from real live\-streaming scenarios, with 10K examples for SFT and 4K interaction tasks for RL\. During evaluation, Qwen3\.6\-35B\-A3B and our trained checkpoints are deployed on one H20, while the Top models are accessed through Bailian API routes\. We evaluate the checkpoints on the offline sets and deployment replay defined in Section[4](https://arxiv.org/html/2608.15763#S4)\. For𝒯1\\mathcal\{T\}\_\{1\}–𝒯3\\mathcal\{T\}\_\{3\}, all model comparisons use the same Harness; for𝒯4\\mathcal\{T\}\_\{4\}, all models are evaluated with the same official IFEval code\.

### 5\.2 Main Results

Table 3:Main results across offline evaluation sets\.𝒯1\\mathcal\{T\}\_\{1\}measures real live\-stream reply quality,𝒯2\\mathcal\{T\}\_\{2\}measures robustness to Harness\-Variant QA,𝒯3\\mathcal\{T\}\_\{3\}measures Tool and Prompt Robustness on synthetic avatar scenarios, and𝒯4\\mathcal\{T\}\_\{4\}measures general instruction following with IFEval\. Acc, Eff, and AVG denote Accuracy, Effectiveness, and their per\-sample average; ADR, RC, TC, LLM, and TCO are the rubric metrics defined in Section[4\.2](https://arxiv.org/html/2608.15763#S4.SS2)\. Fixed\-Harness SFT denotes conventional distilled supervised fine\-tuning without HSA, while HAT \(Ours\) denotes the model obtained by the full HSA\-SFT \+ General OPD \+ HSA\-RL pipeline\.Panel A· Live\-Stream QA quality and general instruction following

Panel B· Broad AI\-generated live\-stream avatar scenarios

Results at a glance▶\\blacktrianglerightFixed\-Harness SFT shortcut\.Fixed\-Harness SFT improves the in\-family task scores but lowers Prompt Robustness \(−\-4\.6\) and IFEval \(−\-7\.7/−\-5\.3\), suggesting that training on one harness version can encourage shortcut learning and reduce robustness and generalization\.▶\\blacktrianglerightFull\-pipeline model\.HAT reaches 94\.8 on𝒯1\\mathcal\{T\}\_\{1\}, exceeding the strongest Top\-model score of 93\.0 on the same business dataset, while avoiding the robustness and generalization drops observed after Fixed\-Harness SFT\.

### 5\.3 Ablation Studies

#### Training Pipeline Ablation

Table[4](https://arxiv.org/html/2608.15763#S5.T4)reports cumulative training through each pathway’s shared initialization, followed by alternative Agentic RL branches under the original Harness and HSA environments\.

Table 4:Training\-pipeline ablation\. Each row adds one stage to the previous checkpoint;↑\\uparrow/↓\\downarrowdenotes a descriptive score change from the row above and does not include retraining variance\.Key insightsFixed\-Harness SFT pattern:Fixed\-Harness SFT improves business\-task scores, but its lower Prompt Robustness and IFEval scores are consistent with a loss of generalization after SFT\. This may arise from catastrophic forgetting\([Luo et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib25)\), shortcut learning\([Geirhos et al\. 2020](https://arxiv.org/html/2608.15763#bib.bib12)\), or memorization of tool\-use patterns rather than robust interpretation of the current Harness context\([Zeng et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib50)\)\.General OPD and HSA\-SFT:General OPD recovers part of the general instruction\-following ability that drops after Fixed\-Harness SFT; HSA\-SFT also avoids the large IFEval drop observed on the non\-augmented path\. Combining the two keeps the model competitive on business quality while maintaining better generalization\.RL initialization matters:Agentic RL obtains a better final operating point when it starts from the HSA\-SFT \+ General OPD checkpoint than when it starts from the Fixed\-Harness SFT \+ General OPD checkpoint, suggesting that RL benefits from a base model with stronger Harness generalization\.HSA\-RL matters:HSA\-RL improves over Agentic RL in the original Harness environment, suggesting that the RL stage should also expose the policy to changing Harness contexts\.

#### Controlled On\-Policy Optimization Comparison

We further examine the effect of OPD in our live\-streaming Harness\. Starting from the checkpoint before Agentic RL, we use the corresponding Agentic\-RL checkpoint as the teacher and distill its behavior through OPD under the same original Harness or HSA environment\.

Table[5](https://arxiv.org/html/2608.15763#S5.T5)shows that OPD transfers part of the Agentic\-RL teacher’s behavior, but its point estimates remain below the paired RL checkpoint\. In the original Harness environment, OPD is lower than RL by 0\.1 points on𝒯1\\mathcal\{T\}\_\{1\}, 0\.2 on IFE\-P, and 4\.8 on Prompt Robustness; in the HSA environment, the corresponding gaps are 0\.9, 1\.8, and 2\.1 points\.

Table 5:Single\-run comparison of Agentic optimization variants\. All configurations share the same Aug\-SFT \+ General OPD initialization;⋆\\starmarks the deployment choice rather than statistical superiority\.Reading\.The RL checkpoints have higher IFEval and Prompt\-Robustness point estimates than their paired OPD checkpoints, while business\-score gaps remain below one point\. The selected HSA\-RL checkpoint is an engineering operating\-point choice; these single\-run contrasts are not method\-level effects\.

#### HSA Stage Isolation

Table[6](https://arxiv.org/html/2608.15763#S5.T6)isolates the effect of HSA at the SFT and RL stages using final checkpoint metrics from runs without General OPD\. In this run, the IFE\-P difference associated with HSA\-SFT is \+9\.4, compared with \+1\.0 for HSA\-RL, suggesting that the main HSA gain comes from the SFT stage\. Figure[5](https://arxiv.org/html/2608.15763#S5.F5)complements this table with reward\-trajectory diagnostics across the four SFT/RL configurations, showing how the corresponding Accuracy, Effectiveness, Skill Selection, and Tool Rationality rewards evolve during optimization\.

Table 6:Single\-run isolation of data augmentation at SFT and RL stages\. The reported differences are descriptive and do not include retraining variance\.

### 5\.4 Controlled Deployment Inference

Table[7](https://arxiv.org/html/2608.15763#S5.T7)reports client\-observed performance in the application’s low\-concurrency operating range\. Every row uses the same fixed 100\-case sample and complete Agent workflow; the API rows traverse vendor\-managed serving routes, while all H20 rows use one local GPU\.

At concurrency 1 and 2, the selected MTP configuration reaches P95 latencies of 8\.114 and 9\.047 s, and all measured requests satisfy the 15\-second bound\. On the same checkpoint and H20, MTP raises decoding throughput from 160\.30 to 271\.40 tokens/s at concurrency 1 \(1\.69×\\times\) and from 129\.22 to 196\.15 at concurrency 2 \(1\.52×\\times\)\. Cross\-checkpoint Base–HAT Decode values are not intrinsic speed comparisons: each policy induces different Agent trajectories and completion\-length distributions, and the arithmetic mean of per\-call ratios is especially sensitive to very short generations\. We therefore use the within\-checkpoint Off–On contrast as the primary acceleration evidence\. Under MTP On, HAT also has higher draft acceptance, so its remaining cross\-checkpoint gap is not attributed to length alone\. The offline quality values elsewhere in this section use MTP Off\. A same\-checkpoint, same\-Judge check on𝒯1\\mathcal\{T\}\_\{1\}records 96\.22 Accuracy and 94\.22 Effectiveness with MTP On, compared with 95\.8 and 93\.8 with MTP Off; thus, no aggregate quality degradation is observed in this run\. The API routes have P95 values above 21 s at concurrency 1, but their hardware, batching, and service load are not observable; we use them only as end\-to\-end operational references\. Appendix[F\.1](https://arxiv.org/html/2608.15763#A6.SS1)reports the CoT\-length and draft\-head diagnostics, quality check, workload shape, stress\-test behavior, and SGLang diagnostics\.

Table 7:Controlled low\-concurrency deployment replay\. Latencies are seconds; Decode is the arithmetic mean of client\-observed completion tokens/s over model calls\. It is length\-sensitive and is not an intrinsic cross\-checkpoint kernel\-speed measure\. TTFT is the first call’s client\-observed P95 and includes network and queueing\. Each row contains 100 measured Agent requests after 10 warm\-up cases\. API rows are operational references, not same\-hardware model comparisons\.
### 5\.5 Held\-Out Evaluation of Harness Edits

We construct an additional held\-out evaluation set for Harness edits by sampling 1,500 cases from real online live\-streaming traffic, with 500 development cases and 1,000 held\-out test cases\. The raw logs contain viewer comments, dialogue history, live\-room context, product information, tool\-use traces, and model responses\. We use an automatic quality check to count bad cases where the model output is not sufficiently natural or colloquial for a live\-streaming host\.

The development split is used only for Harness diagnosis and editing\. We analyze the bad cases on the 500\-case development set and revise editable Harness components, primarily the system prompt and Skills, while keeping the model weights fixed\. The resulting Harness edits are fixed and applied unchanged to both the Fixed\-Harness SFT checkpoint and the HAT checkpoint\. The test split is held out from this editing loop and is used to compare each model before and after applying the same Harness edits\.

Table[8](https://arxiv.org/html/2608.15763#S5.T8.fig1)reports the held\-out test results\. Counts are total detected errors, so lower is better\. The Harness edits reduce the total error count for both the Fixed\-Harness SFT checkpoint and the HAT checkpoint, with the largest absolute reduction observed for HAT\.

Table 8:Harness\-Aware Benchmark held\-out test results before and after Harness edits\. Counts are total detected errors; error reduction is computed relative to the before\-edit count\.
### 5\.6 Human Blind\-Test Protocol and Results

We conduct a human blind test to compare the modular Harness runtime with a ReAct\-style baseline on real live\-streaming requests\. We draw a stratified 100\-request blind\-test set with a fixed random seed so that its scenario composition follows the full pool\.

For each request, both systems run independently on the same input\. The record preserves the user input, dialogue history, complete reasoning and tool\-use trace, and final response from both systems\. During annotation, the two responses are randomly assigned to Model A and Model B, and the annotator is not shown the system identity\. The annotator chooses one of three labels: Harness better, ReAct better, or tie\. Non\-tie decisions require a short reason\. After annotation, all 100 examples are reviewed for data consistency and labeling consistency\.

Table 9:Human blind\-test preference distribution on 100 paired real live\-streaming requests\.Table[9](https://arxiv.org/html/2608.15763#S5.T9.fig1)reports the preference distribution\. Among the 36 examples with a clear preference, Harness wins 35, corresponding to a 97\.2% non\-tie win rate\. A two\-sided binomial test against equal preference givesp<0\.001p<0\.001\. The only ReAct\-preferred example is an ambiguous short comment, where Harness matches an irrelevant FAQ\-style response while ReAct gives a safer transitional reply\. We therefore treat this case as an isolated fallback\-policy failure rather than a recurring disadvantage\.

Table 10:Attribution of the 35 Harness\-preferred examples\. Categories are assigned from annotator rationales after the blind decision\.Table[10](https://arxiv.org/html/2608.15763#S5.T10.fig1)summarizes the attribution of Harness wins\. The largest sources of improvement are input understanding, output reliability, and scenario\-appropriate behavior, which together account for 80\.0% of the Harness\-preferred cases\. These categories include correctly separating live\-room events from viewer questions, avoiding unsupported claims or false promises, handling low\-information comments without forced product promotion, and using available tool evidence more appropriately\. The blind\-test result is consistent with the automatic Judge evaluation and provides an additional human preference check on real live\-streaming requests\.

### 5\.7 Agentic Optimization Diagnostics

Figure 5:Checkpoint trajectories across four training configurations formed by crossing HSA at the SFT and RL stages, evaluated on Accuracy, Effectiveness, Skill Selection, and Tool Rationality\. In these trajectories, HSA\-SFT is associated with higher reply\-quality rewards, while HSA\-RL is associated with stronger agentic\-behavior rewards, especially Skill Selection and Tool Rationality; combining HSA\-SFT and HSA\-RL gives the most balanced late\-stage reward profile\.Figure[5](https://arxiv.org/html/2608.15763#S5.F5)provides a diagnostic view of the four HSA configurations\. The trajectories suggest that HSA\-SFT primarily improves the reply\-quality dimensions, while HSA\-RL raises the agentic\-behavior dimensions, especially Skill Selection and Tool Rationality\. Overall, combining HSA\-SFT and HSA\-RL gives the most balanced late\-stage profile across all four rewards\.

## 6 Deployment and Production Lessons

A production live\-streaming policy must simultaneously keep end\-to\-end P95 below 15 seconds, retain open\-ended instruction following after domain training, and accommodate weekly changes to Skills, tools, and prompts\. The selected HAT checkpoint meets the latency requirement in the controlled low\-concurrency replay, while the reported Fixed\-Harness SFT checkpoint loses general\-set and prompt\-robustness points\. HAT and Harness Evolution are therefore operated as two independently versioned paths\.

### 6\.1 Evolution\-Aware Release Process

Policy checkpoints change through HAT, whereas Skills, prompts, Hooks, and tools change through Harness Evolution \(Section[2\.3](https://arxiv.org/html/2608.15763#S2.SS3)\)\. A release manifest binds the policy checkpoint, Harness commit, tool schemas, fallback thresholds, and evaluation\-set versions\. Every proposed change runs the business regression suite and the general\-capability gate before traffic promotion\. Public\-channel changes receive stricter review because one spoken response reaches the whole room\. The non\-monotonic trajectory in Figure[2](https://arxiv.org/html/2608.15763#S2.F2)motivates an explicit early\-stop rule: when long\-tail fixes begin creating cross\-category regressions, operators retain the last passing version rather than accumulating more rules\.

Three production gatesPromote a version only when it passes the 15\-second latency bound, the business regression suite, and the general\-capability gate\. Policy and Harness versions have independent switches, allowing a regression to be attributed and reversed without rolling back both layers\.

### 6\.2 Failure Containment and Recovery

The runtime uses defense in depth rather than relying on the policy alone\. Parameter Hooks reject malformed tool calls before execution; tool\-result Hooks convert service errors into typed observations; the Skill\-loading and output\-format Hooks force a bounded retry; and the loop terminates after four model rounds\. A first\-call timeout of 3 s, later\-call timeout of 2 s, and 10 s trajectory budget trigger the compact fallback route\. If recovery still fails, the system returns a conservative response or suppresses the public reply instead of inventing product facts\.

Three recurrent high\-severity errors shaped release checks: unsupported store or product claims, false claims that a human is driving the avatar, and tool calls that repeatedly fail or omit required parameters\. Product, price, inventory, and promotion statements must be grounded in the corresponding tool result\. An unsupported\-claim Hook blocks unverifiable attributes, guarantees, discounts, and delivery promises; identity rules require disclosure that the host is an AI avatar\. Guardrail hits, retries, fallbacks, and suppressed replies are logged with the model and Harness versions for incident review\.

### 6\.3 Low\-Concurrency Serving Policy

The serving configuration uses the selected task\-trained policy on one H20 with MTP enabled\. The application is provisioned for one or two simultaneous requests per replica rather than high\-throughput batch inference\. A separate SGLang operational trace is consistent with this workload shape: sampled decode log intervals have mean 2\.03 and median 1 running request, with a P95 of 7 and no queued requests in the captured window\. These are log\-sample\-weighted scheduler observations, not a request\-weighted production concurrency distribution\.

The controlled replay therefore treats concurrency 1–2 as the decision\-relevant range\. In that range, MTP gives the selected checkpoint 1\.69×\\times/1\.52×\\timeshigher client\-observed decoding throughput while maintaining 8\.114/9\.047 s P95 and 100% 15\-second attainment\. The concurrency 4–8 runs are stress tests: the selected checkpoint’s decoding benefit falls to 1\.27×\\times/1\.19×\\times, and MTP no longer improves P95\. This boundary motivates provisioning additional replicas rather than treating MTP as a universal high\-concurrency throughput optimization\.

### 6\.4 Preliminary Online A/B Test on Taobao Live

We deployed the Harness\-based system in Taobao Live’s production digital\-avatar business and conducted an ongoing user\-level online A/B test against the existing ReAct system\. Unlike the controlled replay above, this experiment measures real production traffic\. Users are assigned through stable user\-level bucketing\. At the seven\-day snapshot, the ReAct control contained 678,598 bucketed users and the Harness treatment contained 75,754, corresponding to approximately 89\.96% and 10\.04% of the experiment population\. The arms compare complete production system versions\.

We define GMV per user as multiday cumulative confirmed\-receipt GMV in RMB divided by the number of bucketed users, and orders per user as completed orders divided by the same denominator\. Table[11](https://arxiv.org/html/2608.15763#S6.T11.fig1)reports only relative changes against the ReAct control because the unequal bucket sizes make raw totals unsuitable for direct comparison\.

Table 11:Preliminary seven\-day online A/B point estimates in Taobao Live’s production digital\-avatar business\. Metrics are relative changes of the Harness system against the ReAct control\.Note:This online A/B test is still ongoing in Taobao Live’s production digital\-avatar business\. Values are unadjusted point estimates from the first seven days and are provided for reference only; confidence intervals and significance tests are not available for this preliminary snapshot\.

At this seven\-day production snapshot, the complete Harness system shows observed preliminary uplifts of 5\.54% in confirmed\-receipt GMV per bucketed user and 0\.80% in completed orders per bucketed user relative to ReAct\. These aggregate point estimates characterize the current full\-system comparison and may change as the experiment continues\.

## 7 Related Work

##### Agent architectures and Harness Evolution

ReAct\([Yao et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib44)\)interleaves reasoning and action in a single prompt; Toolformer\([Schick et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib32)\)teaches models to invoke external tools; Reflexion\([Shinn et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib35)\)adds within\-episode self\-correction\. TPTU\-v2\([Kong et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib16)\)improves task planning and tool use in an industrial agent system, while DSPy\([Khattab et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib15)\)compiles declarative modules into optimized prompts\. RIVA\([Abuzakuk et al\. 2026](https://arxiv.org/html/2608.15763#bib.bib1)\)addresses configuration\-drift detection in infrastructure through cross\-validating agents; our problem is complementary because the configuration being changed is the agent’s own execution environment\. Harness Evolution makes that editable runtime—Skills, prompt assembly, Hooks, and tools—an explicit release process and a training\-distribution object\.

##### Prompt, schema, and agent\-evaluation robustness

LLM behavior can vary under semantically equivalent prompt templates\([Zhuo et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib56)\), and tool selection is sensitive to edits in natural\-language tool descriptions\([Faghih et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib11)\)\. SMART\([Qian et al\. 2025b](https://arxiv.org/html/2608.15763#bib.bib30)\)further shows that agents must learn when not to invoke a tool\. These findings motivate varying prompt and tool interfaces rather than evaluating a single frozen surface form\. Stateful benchmarks such asτ\\tau\-bench\([Yao et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib45)\)and ToolSandbox\([Lu et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib23)\)evaluate multi\-turn interaction and intermediate state; ECom\-Bench\([Wang et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib39)\)adds dynamic user simulation for e\-commerce customer support\. Our𝒯3\\mathcal\{T\}\_\{3\}instead targets broad live\-stream avatar interactions and complements the controlled Harness variants in𝒯2\\mathcal\{T\}\_\{2\}\.

##### Agent post\-training

SFT\-based agent training constructs trajectory data for instruction following and tool use\([Zeng et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib49);[Chen et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib5);[Chen et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib7)\); rejection\-sampling and self\-taught variants filter successful rollouts\([Zelikman et al\. 2022](https://arxiv.org/html/2608.15763#bib.bib48);[Yuan et al\. 2023](https://arxiv.org/html/2608.15763#bib.bib47)\)\. RL methods optimize policies through environment interaction\([Putta et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib28);[Xi et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib42)\); ToolRL\([Qian et al\. 2025a](https://arxiv.org/html/2608.15763#bib.bib29)\)studies structured reward design for tool selection and application, while GRPO\([Shao et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib34)\)provides the group\-relative objective used here\. These agent RL approaches build on the reinforcement learning from human feedback \(RLHF\) paradigm\([Ouyang et al\. 2022](https://arxiv.org/html/2608.15763#bib.bib27)\)that first demonstrated effective policy optimization for instruction\-following language models\. RandomWorld\([Sullivan et al\. 2025](https://arxiv.org/html/2608.15763#bib.bib36)\)procedurally generates interactive tools and compositional tasks\. Our environment keeps the real business ontology and randomizes the editable Harness around it\. On\-policy distillation\([Agarwal et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib2)\)anchors general behavior after domain SFT\. Unlike work whose primary novelty is a reward, credit\-assignment rule, or optimizer, our contribution lies in formulating the execution environment as a changing training\-distribution object and evaluating a controlled combination of established mechanisms in production\.

##### Training under distribution shift

Domain adaptation aligns source and target distributions\([Wang et al\. 2022](https://arxiv.org/html/2608.15763#bib.bib40)\), continual learning handles sequential tasks without forgetting\([Wang et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib41)\), and test\-time adaptation adjusts models at inference\([Liang et al\. 2024](https://arxiv.org/html/2608.15763#bib.bib19)\)\. Recent agent\-specific test\-time adaptation learns environment format and dynamics through deployment\-time interaction\([Chen et al\. 2026a](https://arxiv.org/html/2608.15763#bib.bib4)\); it is a complementary option when changing weights or probing the environment online is acceptable\. Harness\-State Augmentation \(HSA\) is a domain\-randomization instantiation\([Tobin et al\. 2017](https://arxiv.org/html/2608.15763#bib.bib38)\): it trains across controlled variation so later conditions within the declared change envelope are more likely to lie within learned support\. The adaptation here is to symbolic, executable, and coupled agent configuration: a valid change may require consistent edits across Skills, prompts, tools, Hooks, and message logic\.𝒯2\\mathcal\{T\}\_\{2\}evaluates in\-family changes, while𝒯3\\mathcal\{T\}\_\{3\}evaluates broader in\-domain scenario transfer; neither constitutes unseen\-family validation\.

## 8 Conclusion

We presented Harness\-Aware Training \(HAT\), a systems\-and\-training methodology for the tension between versioned Harness change and policy stability in live\-streaming digital\-avatar agents\. It treats Harness\-state variation as a training\-distribution design problem within an explicitly defined change envelope\.

Our work makes four contributions\. First, Harness Evolution versions runtime changes separately from policy updates through modular Skills, Hooks, prompts, and tools, and exposes the training problem created by a moving execution environment\. Second, HAT treats Harness\-state variation as part of the training distribution through HSA and a three\-stage pipeline with HSA\-SFT, General OPD, and HSA\-RL\. Third, scoped offline sets, a human\-aligned evaluation panel, and a controlled complete\-Agent deployment replay separate checkpoint quality evidence from serving\-route behavior\. Fourth, the ablations provide practical training insights: General OPD mainly recovers instruction\-following ability after Fixed\-Harness SFT, while HSA is most effective when introduced during SFT\.

In our study, the HAT\-trained compact model reaches 94\.8 Live\-Stream QA AVG and does not exhibit the general\-set drop observed after Fixed\-Harness SFT\. On one H20 with MTP enabled, the controlled replay reaches 8\.114 s P95 at concurrency 1 and 9\.047 s at concurrency 2, with 100% 15\-second attainment; MTP increases client\-observed decoding throughput by 1\.69×\\times/1\.52×\\timesover the same checkpoint without MTP\. Overall, these results show that HAT can produce a latency\-feasible compact agent while preserving the adaptability needed for evolving production Harnesses\.

## Author Contributions

Project Leader:Yuhan Sun Core Contributors:Wenhao Lin, Yongdong Luo, Yibo Hu, Junfeng Ma, Meiguang Jin Contributors:Weihang Pan, Jiaxin Zhao, Zulong Chen

## Acknowledgments

We thank the Alibaba ROLL Team for their invaluable support on training infrastructure\. We are grateful to Dr\. Weihang Pan from Zhejiang University and Zulong Chen from Alibaba Group for their highly valuable contributions and assistance\.

## References

- Abuzakuk et al\. \[2026\]Sami Abuzakuk, Lucas Crijns, Anne\-Marie Kermarrec, Rafael Pires, and Martijn de Vos\.RIVA: Leveraging LLM agents for reliable configuration drift detection\.In*Proceedings of the Sixth European Workshop on Machine Learning and Systems*, 2026\.doi:10\.1145/3805621\.3807644\.
- Agarwal et al\. \[2024\]Rishabh Agarwal, Nino Vieillard, Piotr Stanber, Marc’Aurelio Ranzato, and Matthieu Geist\.On\-policy distillation of language models: Learning from self\-generated mistakes\.In*International Conference on Learning Representations \(ICLR\)*, 2024\.
- Allal\-Chérif et al\. \[2024\]Oihab Allal\-Chérif, Virginia Simón\-Moya, and Ana Cristina Conejero Ballester\.Intelligent influencer marketing: How AI\-powered virtual influencers outperform human influencers\.*Technological Forecasting and Social Change*, 200, 2024\.
- Chen et al\. \[2026a\]Arthur Chen, Zuxin Liu, Jianguo Zhang, Akshara Prabhakar, Zhiwei Liu, Shelby Heinecke, Silvio Savarese, Victor Zhong, and Caiming Xiong\.Test\-time adaptation for LLM agents via environment interaction\.In*International Conference on Learning Representations*, 2026a\.URL[https://openreview\.net/forum?id=OH4PE0TDo0](https://openreview.net/forum?id=OH4PE0TDo0)\.
- Chen et al\. \[2023\]Baian Chen, Chang Shu, Ehsan Shareghi, Nigel Collier, Karthik Narasimhan, and Shunyu Yao\.Fireact: Toward language agent fine\-tuning\.*arXiv preprint arXiv:2310\.05915*, 2023\.
- Chen et al\. \[2026b\]Jian Chen, Yesheng Liang, and Zhijian Liu\.DFlash: Block diffusion for flash speculative decoding\.In*Proceedings of the 43rd International Conference on Machine Learning \(ICML\)*, 2026b\.
- Chen et al\. \[2024\]Zehui Chen, Kuikun Du, Wenqiang Chen, Enze Liu, Terry Yue Zhuo, Jian Lu, and Jie Wang\.Agent\-flan: Designing data and methods of effective agent tuning for large language models\.*arXiv preprint arXiv:2403\.12881*, 2024\.
- Cheng et al\. \[2026\]Xin Cheng, Xingkai Yu, Chenze Shao, Jiashi Li, Yunfan Xiong, Yi Qian, et al\.DSpark: Confidence\-scheduled speculative decoding with semi\-autoregressive generation\.*arXiv preprint arXiv:2607\.05147*, 2026\.
- Cui and Liu \[2023\]Liang Cui and Jing Liu\.Virtual human: A comprehensive survey on academic and applications\.*IEEE Access*, 11:123830–123845, 2023\.
- DeepSeek\-AI \[2024\]DeepSeek\-AI\.Deepseek\-v3 technical report\.*arXiv preprint arXiv:2412\.19437*, 2024\.
- Faghih et al\. \[2025\]Kazem Faghih, Wenxiao Wang, Yize Cheng, Siddhant Bharti, Gaurang Sriramanan, Sriram Balasubramanian, Parsa Hosseini, and Soheil Feizi\.Tool preferences in agentic LLMs are unreliable\.In*Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing*, pages 20954–20969, 2025\.doi:10\.18653/v1/2025\.emnlp\-main\.1060\.
- Geirhos et al\. \[2020\]Robert Geirhos, Jörn\-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann\.Shortcut learning in deep neural networks\.*Nature Machine Intelligence*, 2\(11\):665–673, 2020\.
- Hebbar et al\. \[2026\]Aditya Hebbar et al\.Sia: Self improving ai\.*arXiv preprint arXiv:2605\.27276*, 2026\.
- Jin et al\. \[2025\]Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han\.Search\-r1: Training llms to reason and leverage search engines with reinforcement learning\.*arXiv preprint arXiv:2503\.09516*, 2025\.
- Khattab et al\. \[2023\]Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T\. Joshi, Hanna Mober, et al\.Dspy: Compiling declarative language model calls into self\-improving pipelines\.*arXiv preprint arXiv:2310\.03714*, 2023\.
- Kong et al\. \[2024\]Yilun Kong, Jingqing Ruan, Yihong Chen, Bin Zhang, Tianpeng Bao, Shiwei Shi, Guo Qing Du, Xiaoru Hu, Hangyu Mao, Ziyue Li, Xingyu Zeng, Rui Zhao, and Xueqian Wang\.TPTU\-v2: Boosting task planning and tool usage of large language model\-based agents in real\-world industry systems\.In*Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track*, pages 371–385, 2024\.doi:10\.18653/v1/2024\.emnlp\-industry\.27\.
- Lambert et al\. \[2025\]Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V\. Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, Yuling Gu, Saumya Malik, Victoria Graf, Jena D\. Hwang, Jiangjiang Yang, Ronan Le Bras, Oyvind Tafjord, Chris Wilhelm, Luca Soldaini, Noah A\. Smith, Yizhong Wang, Pradeep Dasigi, and Hannaneh Hajishirzi\.Tulu 3: Pushing frontiers in open language model post\-training\.*arXiv preprint arXiv:2411\.15124*, 2025\.
- Li et al\. \[2024\]Yuhui Li, Fangyun Wei, Chao Zhang, and Hongyang Zhang\.EAGLE: Speculative sampling requires rethinking feature uncertainty\.In*Proceedings of the 41st International Conference on Machine Learning \(ICML\)*, 2024\.
- Liang et al\. \[2024\]Jian Liang, Ran He, and Tieniu Tan\.A comprehensive survey on test\-time adaptation under distribution shifts\.*International Journal of Computer Vision*, 2024\.
- Lin et al\. \[2026\]Zijie Lin et al\.Agentic harness engineering\.*arXiv preprint arXiv:2604\.25850*, 2026\.
- Liu et al\. \[2026\]Shih\-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, Peter Belcak, Mingjie Liu, Min\-Hung Chen, Hongxu Yin, Yu\-Chiang Frank Wang, Kwang\-Ting Cheng, Yejin Choi, Jan Kautz, and Pavlo Molchanov\.GDPO: Group reward\-decoupled normalization policy optimization for multi\-reward RL optimization\.*arXiv preprint arXiv:2601\.05242*, 2026\.
- Liu et al\. \[2025\]Yifan Liu, Lin Wang, Shuo Yang, and Yining Wang\.AI\-powered digital streamers for online retail: Empirical evidence and design strategies through experiments\.*Information Systems Research*, 36\(4\), 2025\.
- Lu et al\. \[2025\]Jiarui Lu, Thomas Holleis, Yizhe Zhang, Bernhard Aumayer, Feng Nan, Felix Bai, Shuang Ma, Shen Ma, Mengyu Li, Guoli Yin, Zirui Wang, and Ruoming Pang\.Toolsandbox: A stateful, conversational, interactive evaluation benchmark for llm tool use capabilities\.In*Findings of the Association for Computational Linguistics: NAACL 2025*, 2025\.
- Lu and Lab \[2025\]Kevin Lu and Thinking Machines Lab\.On\-policy distillation\.*Thinking Machines Lab: Connectionism*, 2025\.doi:10\.64434/tml\.20251026\.https://thinkingmachines\.ai/blog/on\-policy\-distillation\.
- Luo et al\. \[2025\]Yun Luo, Zhen Yang, Fandong Meng, Yafu Li, Jie Zhou, and Yue Zhang\.An empirical study of catastrophic forgetting in large language models during continual fine\-tuning\.*IEEE Transactions on Audio, Speech and Language Processing*, 2025\.
- Ning et al\. \[2025\]Xuying Ning, Katherine Tieu, Dongqi Fu, et al\.Code as agent harness: Toward executable, verifiable, and stateful agent systems\.*arXiv preprint*, 2025\.
- Ouyang et al\. \[2022\]Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al\.Training language models to follow instructions with human feedback\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2022\.
- Putta et al\. \[2024\]Pranav Putta, Edmund Mills, Naman Garg, Sumeet Motwani, Chelsea Finn, Divyansh Garg, and Rafael Rafailov\.Agentq: Advanced reasoning and learning for autonomous ai agents\.*arXiv preprint arXiv:2408\.07199*, 2024\.
- Qian et al\. \[2025a\]Cheng Qian, Emre Can Acikgoz, Qi He, Hongru Wang, Xiusi Chen, Dilek Hakkani\-Tur, Gokhan Tur, and Heng Ji\.ToolRL: Reward is all tool learning needs\.In*Advances in Neural Information Processing Systems*, volume 38, 2025a\.
- Qian et al\. \[2025b\]Cheng Qian, Emre Can Acikgoz, Hongru Wang, Xiusi Chen, Avirup Sil, Dilek Hakkani\-Tur, Gokhan Tur, and Heng Ji\.SMART: Self\-aware agent for tool overuse mitigation\.In*Findings of the Association for Computational Linguistics: ACL 2025*, pages 4604–4621, 2025b\.doi:10\.18653/v1/2025\.findings\-acl\.239\.
- Qu et al\. \[2026\]Wanying Qu, Qinghua Mao, Yu Li, et al\.She: Trajectory\-driven safety harness evolution for llm agents\.*arXiv preprint arXiv:2608\.xxxxx*, 2026\.
- Schick et al\. \[2024\]Timo Schick, Jane Dwivedi\-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom\.Toolformer: Language models can teach themselves to use tools\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2024\.
- Schulman et al\. \[2017\]John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov\.Proximal policy optimization algorithms\.*arXiv preprint arXiv:1707\.06347*, 2017\.
- Shao et al\. \[2024\]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y\. K\. Li, Y\. Wu, et al\.Deepseekmath: Pushing the limits of mathematical reasoning in open language models\.*arXiv preprint arXiv:2402\.03300*, 2024\.
- Shinn et al\. \[2023\]Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao\.Reflexion: Language agents with verbal reinforcement learning\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2023\.
- Sullivan et al\. \[2025\]Michael Sullivan, Mareike Hartmann, and Alexander Koller\.Procedural environment generation for tool\-use agents\.In*Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing*, pages 18544–18562, 2025\.doi:10\.18653/v1/2025\.emnlp\-main\.936\.
- Sun et al\. \[2025\]Yuhan Sun, Zixuan Huang, Wenhao Cui, Sicheng Xiong, Yuxuan Guo, Mingshuo Jin, and Jiaxin Ma\.Livethinking: Enabling real\-time efficient reasoning for AI\-powered livestreaming via reinforcement learning\.*arXiv preprint arXiv:2510\.07685*, 2025\.
- Tobin et al\. \[2017\]Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel\.Domain randomization for transferring deep neural networks from simulation to the real world\.*arXiv preprint arXiv:1703\.06907*, 2017\.
- Wang et al\. \[2025\]Haoxin Wang, Xianhan Peng, Xucheng Huang, Yizhe Huang, Ming Gong, Chenghan Yang, Yang Liu, and Ling Jiang\.ECom\-Bench: Can LLM agent resolve real\-world e\-commerce customer support issues?In*Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track*, pages 276–284, 2025\.doi:10\.18653/v1/2025\.emnlp\-industry\.19\.
- Wang et al\. \[2022\]Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip S\. Yu\.Generalizing to unseen domains: A survey on domain generalization\.*IEEE Transactions on Knowledge and Data Engineering*, 2022\.
- Wang et al\. \[2024\]Liyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu\.A comprehensive survey of continual learning: Theory, method and application\.*IEEE Transactions on Pattern Analysis and Machine Intelligence*, 2024\.
- Xi et al\. \[2024\]Zhiheng Xi, Yiwen Zheng, Xingjian Tian, Wei Chen, Tao Gui, Qi Zhang, and Xuanjing Huang\.Agentgym: Evolving large language model\-based agents across diverse environments\.*arXiv preprint arXiv:2406\.04151*, 2024\.
- Yang et al\. \[2025\]Fang\-Chun Yang, Paola Acevedo, Shiqing Guo, Myounghoon Choi, and Christos Mousas\.Embodied conversational agents in extended reality: A systematic review\.*IEEE Access*, 2025\.
- Yao et al\. \[2023\]Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao\.React: Synergizing reasoning and acting in language models\.In*International Conference on Learning Representations \(ICLR\)*, 2023\.
- Yao et al\. \[2024\]Shunyu Yao, Noah Shinn, and Karthik Narasimhan\.τ\\tau\-bench: A benchmark for tool\-agent\-user interaction in real\-world domains\.*arXiv preprint arXiv:2406\.12045*, 2024\.
- Yu et al\. \[2025\]Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Weinan Dai, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen, Chengyi Wang, Hongli Yu, Yuxuan Song, Xiangpeng Wei, Hao Zhou, Jingjing Liu, Wei\-Ying Ma, Ya\-Qin Zhang, Lin Yan, Mu Qiao, Yonghui Wu, and Mingxuan Wang\.DAPO: An open\-source LLM reinforcement learning system at scale\.*arXiv preprint arXiv:2503\.14476*, 2025\.
- Yuan et al\. \[2023\]Zheng Yuan, Hongyi Yuan, Chengpeng Li, Guanting Dong, Keming Lu, Chuanqi Tan, Chang Zhou, and Jingren Zhou\.Scaling relationship on learning mathematical reasoning with large language models\.*arXiv preprint arXiv:2308\.01825*, 2023\.
- Zelikman et al\. \[2022\]Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D\. Goodman\.Star: Bootstrapping reasoning with reasoning\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2022\.
- Zeng et al\. \[2024\]Aohan Zeng, Mingdao Liu, Rui Lu, Bowen Wang, Xiao Liu, Yuxiao Dong, and Jie Tang\.Agenttuning: Enabling generalized agent abilities for llms\.*arXiv preprint arXiv:2310\.12823*, 2024\.
- Zeng et al\. \[2025\]Yirong Zeng, Xiao Ding, Yutai Hou, Yuxian Wang, Li Du, Juyi Dai, Qiuyang Ding, Duyu Tang, Dandan Tu, Weiwen Liu, et al\.Tool zero: Training tool\-augmented llms via pure rl from scratch\.In*EMNLP \(Findings\)*, pages 9135–9147, 2025\.
- Zhang et al\. \[2026\]Zhenyu Zhang et al\.Self\-harness\.*arXiv preprint arXiv:2606\.09498*, 2026\.
- Zheng et al\. \[2025\]Chujie Zheng, Shixuan Liu, Mingze Li, Xiong\-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, Jingren Zhou, and Junyang Lin\.Group sequence policy optimization\.*arXiv preprint arXiv:2507\.18071*, 2025\.URL[https://arxiv\.org/abs/2507\.18071](https://arxiv.org/abs/2507.18071)\.
- Zheng et al\. \[2024\]Lianmin Zheng, Wei\-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P\. Xing, Hao Zhang, Joseph E\. Gonzalez, and Ion Stoica\.Judging llm\-as\-a\-judge with mt\-bench and chatbot arena\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2024\.
- Zhou et al\. \[2023\]Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou\.Instruction\-following evaluation for large language models\.*arXiv preprint arXiv:2311\.07911*, 2023\.
- Zhou et al\. \[2025\]Yuhang Zhou et al\.Gigpo: Group\-in\-group policy optimization for llm agent training\.In*Advances in Neural Information Processing Systems \(NeurIPS\)*, 2025\.
- Zhuo et al\. \[2024\]Jingming Zhuo, Songyang Zhang, Xinyu Fang, Haodong Duan, Dahua Lin, and Kai Chen\.ProSA: Assessing and understanding the prompt sensitivity of LLMs\.In*Findings of the Association for Computational Linguistics: EMNLP 2024*, pages 1950–1976, 2024\.doi:10\.18653/v1/2024\.findings\-emnlp\.108\.

## Appendix ARuntime Interface and Skill Inventories

Table 12:Working inventory of built\-in and MCP\-discovered runtime interfaces\.Table 13:Working Skill inventory and routing responsibilities\.
## Appendix BHarness Evolution Protocol and Diagnostics

### B\.1 Evolution Loop

Each Harness Evolution stage holds model fixed and changes only editable runtime modules\. The evolution loop contains five steps:

1. 1\.AI diagnosis\.Analyze the latest evaluation result, cluster the top failure causes, attach representative bad cases, and propose root causes\.
2. 2\.Human confirmation\.A developer reviews the proposed skill, prompt, hook, or tool changes, then accepts, modifies, or rejects the plan\.
3. 3\.AI\-assisted editing\.Apply the confirmed changes to skill definitions, prompt assembly, Hooks, and tool logic while preserving a versioned configuration snapshot\.
4. 4\.Human\-triggered evaluation\.Run inference and scoring on the dev\-set\.
5. 5\.Regression check and planning\.Compare paired category changes, inspect new regressions, and let a human choose promotion, rollback, another evolution stage, or stopping\.

The stopping rule is operational: stop when the remaining errors are sparse long\-tail cases and local rule additions are more likely to cause cross\-category regression than systematic improvement\.

### B\.2 What Changed across Evolution Stages

Table 14:Fixed\-policy Harness Evolution record\. Values come from six direct evaluation summaries on the same 482\-item dev\-set\. “Selected” is an engineering early\-stop decision\.The process report attributes the largest Accuracy gains to negative\-feedback chat \(\+60\.7 points\), after\-sales and returns \(\+53\.3\), and purchase/payment cases \(\+33\.3\)\. The largest Effectiveness gains occur for emotion expression \(\+27\.3\), membership or gift benefits \(\+18\.8\), and positive\-feedback or already\-purchased cases \(\+14\.3\)\. These category values are engineering diagnostics rather than pre\-specified inferential analyses\.

After Evolution 2, the report records 36 remaining Accuracy failures across 16 categories and 35 Effectiveness failures without a dominant systematic cluster\. Evolutions 3–4 show why “more rules” is not equivalent to better evolution: targeted long\-tail fixes create interactions across Skills, Hooks, and global instructions\. This observation motivates configuration diversity during training and regression\-aware early stopping during production evolution\.

## Appendix CHSA\-RL Training Details

Algorithm[1](https://arxiv.org/html/2608.15763#alg1)gives the complete HSA\-RL training procedure that combines HSA with GDPO multi\-dimensional advantage estimation, the auxiliary CoT\-length dimension, GSPO sequence\-level importance sampling, and zero\-advantage sample filtering\. Notation follows Section[3](https://arxiv.org/html/2608.15763#S3):GGis the group size;𝒮=\{tool,reply\}\\mathcal\{S\}=\\\{\\texttt\{tool\},\\texttt\{reply\}\\\};𝒟tool=\{Tool Rationality,Skill Selection\}\\mathcal\{D\}\_\{\\texttt\{tool\}\}=\\\{\\text\{Tool Rationality\},\\text\{Skill Selection\}\\\};𝒟reply=\{Accuracy,Effectiveness\}\\mathcal\{D\}\_\{\\texttt\{reply\}\}=\\\{\\text\{Accuracy\},\\text\{Effectiveness\}\\\};λ=0\.1\\lambda=0\.1is the post\-normalization CoT weight; andε\\varepsilonis the PPO clip range\.

Figure[6](https://arxiv.org/html/2608.15763#A3.F6)records the corresponding single\-run training diagnostics\. The first\-to\-last 100\-step window means move from 290\.6 to 171\.2 extracted tokens for tool\-call CoT and from 93\.6 to 37\.0 for final\-reply CoT\. The logged mean penalty scoreq⁡\(L\)q\(L\)moves from 0\.466 to 0\.272, equivalent to a raw reward contribution of−0\.0466\-0\.0466to−0\.0272\-0\.0272before GDPO normalization\. The strongest reduction occurs early, followed by a noisy plateau\. These curves show shorter reasoning traces in this run\.

Figure 6:Single\-run CoT\-length diagnostics for the selected HSA\-RL trajectory\. Faint lines show recorded step values and bold lines show 25\-step trailing means\. Panel \(a\) separates tool\-call and final\-reply CoT lengths and marks the no\-penalty \(L≤100L\\leq 100\), linear \(100<L<200100<L<200\), and saturated \(L≥200L\\geq 200\) regions\. Panel \(b\) shows the logged scoreq⁡\(L\)∈\[0,1\]q\(L\)\\in\[0,1\]and the equivalent raw reward contribution−0\.1​q​\(L\)\-0\.1q\(L\)\. The figure is a training diagnostic, not a causal ablation of latency or quality\.Algorithm 1Agentic RL with Harness\-State Augmentation \(HSA\-RL\)1:Policy

πθ\\pi\_\{\\theta\}; group size

GG; state groups

𝒮=\{tool,reply\}\\mathcal\{S\}=\\\{\\texttt\{tool\},\\texttt\{reply\}\\\}; reward dims

𝒟tool\\mathcal\{D\}\_\{\\texttt\{tool\}\}and

𝒟reply\\mathcal\{D\}\_\{\\texttt\{reply\}\}; CoT weight

λ\\lambda; clip range

ε\\varepsilon
2:foreach training iterationdo

3:Sample a mini\-batch of queries and augmented Harness states

\(x,h\)\(x,h\)
4:

πθold←πθ\\pi\_\{\\theta\_\{\\mathrm\{old\}\}\}\\leftarrow\\pi\_\{\\theta\}; initialize

ℒtool←0\\mathcal\{L\}\_\{\\texttt\{tool\}\}\\leftarrow 0,

ℒreply←0\\mathcal\{L\}\_\{\\texttt\{reply\}\}\\leftarrow 0
5:foreach

\(x,h\)\(x,h\)in the mini\-batchdo

6:for

i=1,…,Gi=1,\\ldots,Gdo

7:Roll out trajectory

τi\\tau\_\{i\}under

πθold\(⋅∣x,h\)\\pi\_\{\\theta\_\{\\mathrm\{old\}\}\}\(\\cdot\\mid x,h\)
8:Split

τi\\tau\_\{i\}into tool segment

yitooly\_\{i\}^\{\\texttt\{tool\}\}and reply segment

yireplyy\_\{i\}^\{\\texttt\{reply\}\}
9:Create masks

mitoolm\_\{i\}^\{\\texttt\{tool\}\}and

mireplym\_\{i\}^\{\\texttt\{reply\}\}over generated action tokens; mask prompts and observations

10:Score

\{ri\(d\)\}d∈𝒟tool\\\{r\_\{i\}^\{\(d\)\}\\\}\_\{d\\in\\mathcal\{D\}\_\{\\texttt\{tool\}\}\}from tool behavior and

\{ri\(d\)\}d∈𝒟reply\\\{r\_\{i\}^\{\(d\)\}\\\}\_\{d\\in\\mathcal\{D\}\_\{\\texttt\{reply\}\}\}from final\-reply quality

11:Extract CoT lengths

LitoolL\_\{i\}^\{\\texttt\{tool\}\}and

LireplyL\_\{i\}^\{\\texttt\{reply\}\}and compute

qis←q⁡\(Lis\)q\_\{i\}^\{s\}\\leftarrow q\(L\_\{i\}^\{s\}\)for

s∈𝒮s\\in\\mathcal\{S\}
12:endfor

13:foreach state group

s∈\{tool,reply\}s\\in\\\{\\texttt\{tool\},\\texttt\{reply\}\\\}do

14:foreach reward dimension

d∈𝒟sd\\in\\mathcal\{D\}\_\{s\}do

15:

A^i\(d\)←\(ri\(d\)−μs\(d\)\)/σs\(d\)\\hat\{A\}\_\{i\}^\{\(d\)\}\\leftarrow\(r\_\{i\}^\{\(d\)\}\-\\mu\_\{s\}^\{\(d\)\}\)/\\sigma\_\{s\}^\{\(d\)\}for

i=1,…,Gi=1,\\ldots,G
16:endfor

17:

A^i\(CoT,s\)←\(qis−μs\(CoT\)\)/σs\(CoT\)\\hat\{A\}\_\{i\}^\{\(\\mathrm\{CoT\},s\)\}\\leftarrow\(q\_\{i\}^\{s\}\-\\mu\_\{s\}^\{\(\\mathrm\{CoT\}\)\}\)/\\sigma\_\{s\}^\{\(\\mathrm\{CoT\}\)\}
18:

A^is←∑d∈𝒟sA^i\(d\)−λ​A^i\(CoT,s\)\\hat\{A\}\_\{i\}^\{s\}\\leftarrow\\sum\_\{d\\in\\mathcal\{D\}\_\{s\}\}\\hat\{A\}\_\{i\}^\{\(d\)\}\-\\lambda\\hat\{A\}\_\{i\}^\{\(\\mathrm\{CoT\},s\)\}; normalize to

A^is,⋆\\hat\{A\}\_\{i\}^\{s,\\star\}
19:foreach sequence

i=1,…,Gi=1,\\ldots,Gwith

A^is,⋆≠0\\hat\{A\}\_\{i\}^\{s,\\star\}\\neq 0do

20:

ρi,ts​\(θ\)←πθ​\(yi,ts∣yi,<ts,x,h\)πθold​\(yi,ts∣yi,<ts,x,h\)\\rho\_\{i,t\}^\{s\}\(\\theta\)\\leftarrow\\dfrac\{\\pi\_\{\\theta\}\(y\_\{i,t\}^\{s\}\\mid y\_\{i,<t\}^\{s\},x,h\)\}\{\\pi\_\{\\theta\_\{\\mathrm\{old\}\}\}\(y\_\{i,t\}^\{s\}\\mid y\_\{i,<t\}^\{s\},x,h\)\}
21:

ℓi,ts←−min⁡\(ρi,ts​A^is,⋆,clip⁡\(ρi,ts,1−ε,1\+ε\)​A^is,⋆\)\\ell\_\{i,t\}^\{s\}\\leftarrow\-\\min\\bigl\(\\rho\_\{i,t\}^\{s\}\\hat\{A\}\_\{i\}^\{s,\\star\},\\mathrm\{clip\}\(\\rho\_\{i,t\}^\{s\},1\{\-\}\\varepsilon,1\{\+\}\\varepsilon\)\\hat\{A\}\_\{i\}^\{s,\\star\}\\bigr\)
22:

ℒs←ℒs\+∑tmi,ts​ℓi,ts∑tmi,ts\\mathcal\{L\}\_\{s\}\\leftarrow\\mathcal\{L\}\_\{s\}\+\\dfrac\{\\sum\_\{t\}m\_\{i,t\}^\{s\}\\ell\_\{i,t\}^\{s\}\}\{\\sum\_\{t\}m\_\{i,t\}^\{s\}\}
23:endfor

24:endfor

25:endfor

26:Update

θ\\thetaby minimizing the mean of

ℒtool\\mathcal\{L\}\_\{\\texttt\{tool\}\}and

ℒreply\\mathcal\{L\}\_\{\\texttt\{reply\}\}
27:endfor

## Appendix DJudge as a Harness Evolution Case

The Final Evaluation Judge is a Harness\-based panel that aggregates DeepSeek\-V4\-pro, GLM\-5\.2, and Qwen3\.7\-max by majority vote\. Its Skills, prompts, and tool\-verification procedures are evolved through disagreement analysis against human labels\. Figure[7](https://arxiv.org/html/2608.15763#A4.F7)shows the trajectory on 482 human\-labeled interactions: Accuracy agreement moves from 81\.54% at Evolution 1 to 90\.46% at Evolution 6, while Effectiveness moves from a 63\.70% uncalibrated baseline to an 83\.40% peak\. The selected multi\-model voting configuration reaches 82\.2% Effectiveness agreement\. Disagreement clusters include missed multi\-comment intents, unsupported capability claims, incorrect product linkage, and inconsistent grading of partial answers; revisions target these recurring classes rather than individual model outputs\.

Figure 7:Judge alignment to a human\-labeled calibration cohort \(n=482n=482\)\. Accuracy and Effectiveness agreement improve through evidence\-tool, rubric, and voting revisions; the cohort evaluates the scoring instrument rather than the policy\.
## Appendix EEvaluation and Reproducibility Details

### E\.1 Judge Roles and Independence

Five roles must not be conflated: the DeepSeek\-V4\-pro SFT teacher generates supervised candidates; the rejection\-sampling Judge filters them; the DeepSeek\-V4\-flash Reward Judge scores HSA\-RL trajectories; the Final Evaluation Judge produces offline scores by majority vote over DeepSeek\-V4\-pro, GLM\-5\.2, and Qwen3\.7\-max; and human labels provide the calibration reference\.

The Reward and Final Evaluation Judges have zero model overlap: DeepSeek\-V4\-flash is not a member of the final three\-model panel\. DeepSeek\-V4\-pro is both SFT teacher and one final voter, but a two\-of\-three majority prevents that vote from deciding any item alone\. GLM\-5\.2 and Qwen3\.7\-max supply two non\-teacher votes, and all three voters can call evidence tools when checking product facts\. Human alignment is measured before the panel is used for model comparison\.

### E\.2 Offline Scoring and Implementation

Accuracy is binary; Effectiveness uses 0, 0\.5, and 1; AVG is the per\-sample mean of the two computed before rounding\.𝒯3\\mathcal\{T\}\_\{3\}reports Tool and Prompt robustness separately, and IFEval uses its official prompt\- and instruction\-level evaluators\. The policy is Qwen3\.6\-35B\-A3B\. HSA\-RL uses GRPO with Accuracy, Effectiveness, Tool Rationality, and Skill Selection rewards under Skill, tool, prompt, Hook, and message\-processing perturbations\. All offline model comparisons use identical set versions and the same Final Evaluation Judge\.

### E\.3𝒯3\\mathcal\{T\}\_\{3\}Robustness Subset Construction

𝒯3\\mathcal\{T\}\_\{3\}contains two synthetic robustness subsets generated from a real\-business scenario seed library: Tool Robustness \(n=1,532n=1\{,\}532\) and Prompt Robustness \(n=491n=491\)\. The seed library is built from live\-streaming e\-commerce scenarios and records the product context, viewer intent, possible tool needs, and reply constraints for each scenario family\.

##### Construction pipeline

We use a shared seed\-driven generation pipeline and instantiate it into two orthogonal branches\. The seed pools are extracted from real live\-streaming business data, including product context, FAQ\-style knowledge, dialogue history, and historical tool\-use traces\. For each item, an LLM generator expands a sampled business seed into a natural viewer comment, live\-room or product context, and the expected behaviors to be evaluated\. The generated item is then converted into the Harness input format with the corresponding tool configuration, instruction configuration, simulated tool\-return map, and evaluation rubrics\. We apply rule\-based and LLM\-based consistency checks to ensure that the query, context, tool returns, instructions, and rubrics are mutually consistent\. Items that fail validation are discarded and regenerated\.

The Tool Robustness subset focuses on tool selection under distractor tools, while the Prompt Robustness subset additionally constructs instruction\-following prompts to test whether the model can satisfy valid prompt constraints\. The corresponding metrics are Tool Called \(TC\), Response Contains \(RC\), and Addresses Core Query \(ADR\) for Tool Robustness, and Addresses Core Query \(ADR\), LLM Judge \(LLM\), and Tool Call Order \(TCO\) for Prompt Robustness\. Each final item may contain multiple rubrics\. AVG is the fraction of passed rubrics computed per item before aggregation\. Overall,𝒯3\\mathcal\{T\}\_\{3\}tests model robustness to unseen in\-domain tools and instructions in synthetic live\-streaming scenarios\.

### E\.4 Independent Bootstrap Stability Check

Table[15](https://arxiv.org/html/2608.15763#A5.T15)reports an independent paired\-bootstrap check on a separate fixed evaluation run with cached model outputs\. The intervals are computed over evaluation items using 10,000 paired bootstrap resamples\. Because this run regenerates model answers rather than reusing the exact outputs underlying the main tables, small differences from the main\-table point estimates arise from generation stochasticity\. These results are therefore used as a stability check for the qualitative contrasts rather than as replacements for the reported main results\.Pboot​\(Δ≤0\)P\_\{\\mathrm\{boot\}\}\(\\Delta\\leq 0\)denotes the fraction of bootstrap replicates whose contrast is non\-positive\.

Table 15:Independent paired\-bootstrap stability check\. Differences are reported in percentage points; intervals are 95% percentile bootstrap intervals over evaluation items\.

## Appendix FDeployment Inference and Operating Details

### F\.1 Deployment Inference Benchmark

#### Protocol and Workload Shape

The deployment set contains 110 cases, with 10 warm\-up cases excluded and 100 requests retained\. The client runs on an office\-network host Harness, real RAG and MCP services, a four\-round Agent limit, and a 5\-second tool timeout\. Generation uses temperature 1\.0, top\-pp0\.95, and a 16,384\-token initial generation limit with configured step\-up to 65,536\. The reported local configurations use one H20, BF16, TP=PP=1, and no quantization\. A separate FP8 trial did not yield a material additional serving gain under this workload, so the reported runs retain BF16\. The model’s NextN MTP module is the draft head; the serving engine consumes its predictions through an EAGLE\-style speculative generation and verification path\. Thus, NextN identifies the model\-side draft module and its weights, whereas EAGLE describes how those draft predictions are used during inference\.

Wall latency starts at client dispatch and ends when the Agent produces its final response\. TTFT is measured to the first streamed chunk of the first model call; it includes client networking, request queueing, and server work, so it is not pure kernel prefill time\. Per\-call Decode TPS is completion tokens divided by call latency minus TTFT, and the table reports the arithmetic mean of these per\-call ratios\. This estimator is sensitive to completion length and timer granularity for very short calls; it is not an intrinsic cross\-checkpoint kernel\-speed measurement\. We do not report client\-derived Prefill TPS, TPOT, or inter\-token latency, and we do not equate this per\-call metric with SGLang’s interval\-level aggregate generation throughput\. Execution success means the Agent returned normally; it does not score response correctness\. Table[18](https://arxiv.org/html/2608.15763#A7.T18)records the workload presented to each C=1 route\. Calls and tokens vary because each model chooses its own Agent trajectory\. API usage supplies token counts for the remote routes; the self\-hosted routes use the serving tokenizer\. Token counts from different tokenizers are useful workload descriptors but are not identical measures of Chinese text length\.

The output distributions explain why cross\-checkpoint mean Decode TPS should be read as workload\-shaped\. At C=1 with MTP disabled, the Base route generates 115 tokens per call on average \(median 78; P95 327\.8\), whereas HAT generates 106 \(median 92\.5; P95 230\.7\)\. Base has a 9\-token call whose recorded post\-TTFT interval is 1 ms, producing a 9,000\-token/s ratio; consequently its mean Decode TPS is 195\.42 while its median is 162\.63\. HAT has mean/median values of 160\.30/159\.71\. We therefore attribute the raw Base–HAT mean gap chiefly to generated\-length and trajectory distributions together with short\-call timing granularity, not to intrinsic decoding speed\. The within\-checkpoint MTP Off–On comparisons remain the primary acceleration evidence\. Under MTP On, the residual cross\-checkpoint difference also coincides with the higher acceptance statistics reported in Table[20](https://arxiv.org/html/2608.15763#A8.T20), so it is not attributed to length alone\.

#### MTP Draft\-Head Adaptation

We compare two standalone MTP\-adaptation trajectories using sequences collected from HSA\-RL policy rollouts\. For the task\-trained trajectory, the policy checkpoint starts fromS0=S\_\{0\}=HSA\-SFT \+ General OPD, while the NextN MTP parameters are transplanted from Qwen3\.6\-35B\-A3B\. For the base trajectory, the policy and its native, factory\-co\-trained NextN head start from the default Qwen3\.6\-35B\-A3B checkpoint\. In both cases, the policy produces rollout labels and the MTP update is performed separately: policy parameters are treated as fixed by the MTP objective, and only the draft\-head parameters are optimized\. Both configurations use the same EAGLE\-style speculative execution path at inference, so the comparison concerns draft\-weight provenance and adaptation, not different speculative algorithms\.

Let a rollout bey1:Ty\_\{1:T\}and letKKbe the number of future\-token offsets\. The MTP loss is the standard multi\-token cross\-entropy that predicts tokent\+1\+kt\+1\+kfrom the prefix throughtt:

ℒMTP​\(ϕ\)=−1Z∑k=0K−1∑t=1T−1−klogpϕ\(k\)\(yt\+1\+k∣y≤t\),Z=∑k=0K−1\(T−1−k\),\\begin\{split\}\\mathcal\{L\}\_\{\\mathrm\{MTP\}\}\(\\phi\)&=\-\\frac\{1\}\{Z\}\\sum\_\{k=0\}^\{K\-1\}\\sum\_\{t=1\}^\{T\-1\-k\}\\log p\_\{\\phi\}^\{\(k\)\}\\\!\\left\(y\_\{t\+1\+k\}\\mid y\_\{\\leq t\}\\right\),\\\\ Z&=\\sum\_\{k=0\}^\{K\-1\}\(T\-1\-k\),\\end\{split\}\(6\)whereϕ\\phidenotes the MTP parameters andk=0k=0is next\-token prediction\. Thus RL supplies on\-policy rollout sequences and labels, while the standalone MTP step uses only supervised multi\-token CE; no reward or advantage enters Equation[6](https://arxiv.org/html/2608.15763#A6.E6)\.

Figure 8:Single\-run standalone MTP\-loss trajectories for a task\-trained policy with a transplanted and subsequently post\-trained NextN head, and for the base model with its native factory\-trained head\. Faint lines show step\-level loss; bold lines show a 25\-step trailing mean\. The runs have different lengths and provide adaptation diagnostics only: they do not compare convergence speed, training stability, sample efficiency, or compute efficiency\.Despite noisy per\-step losses, the grafted\-head trajectory moves from a higher early\-loss region toward the late\-training range of the native\-head trajectory\. This supports the practical conclusion that a task\-trained policy can accept a base\-model NextN head and adapt it through subsequent standalone training on policy\-rollout labels\. In an unplotted pilot, omitting the base MTP initialization produced a very high initial loss\. We also observed that updating the main policy while leaving the MTP head untrained reduced draft acceptance to zero, consistent with the stale draft weights no longer predicting the updated policy\. These are run\-specific engineering observations rather than claims of universal or multi\-seed convergence\.

#### MTP across Concurrency

Table[17](https://arxiv.org/html/2608.15763#A6.T17)separates the deployment operating range \(C=1–2\) from stress conditions \(C=4–8\)\. At C=1–2, MTP increases the selected checkpoint’s Decode TPS by 1\.69×\\times/1\.52×\\timeswhile retaining 100% 15\-second attainment\. At C=4–8, the speedup falls to 1\.27×\\times/1\.19×\\timesand P95 no longer improves\. The base model shows the same qualitative saturation boundary with a smaller low\-concurrency gain\. We therefore treat MTP as a low\-concurrency latency optimization, not as a universal capacity multiplier\.

Table 16:Same\-checkpoint𝒯1\\mathcal\{T\}\_\{1\}quality check with MTP Off and On \(n=978n=978\)\. Both rows use the same Final Evaluation Judge\. AVG is the arithmetic mean of Accuracy and Effectiveness\.Table 17:MTP trade\-off across concurrency\. Arrows show Off→\\rightarrowOn\. Decode TPS is client\-observed mean per call; Wall and TTFT are P95 seconds\. C=1, 2, and 8 are single runs\. C=4 values are means of three run\-level statistics, not pooled percentiles\.C=4 run ranges \(Off; On\)\.Qwen3\.6 Base—Decode: 102\.13–107\.33; 111\.11–116\.49\. Wall P95: 12\.931–13\.756; 14\.957–18\.054\. TTFT P95: 0\.563–0\.758; 0\.942–1\.226\. Attainment: 96–98%; 89–95%\. Ours—Decode: 97\.40–102\.55; 121\.61–130\.18\. Wall P95: 10\.269–11\.606; 10\.532–14\.205\. TTFT P95: 0\.587–0\.784; 1\.305–1\.413\. Attainment: 99–100%; 96–100%\.

## Appendix GDeployment Workload Shape

Table 18:C=1 workload shape\. Values are means over successful complete\-Agent requests or their constituent model calls\.
## Appendix HSupplementary Reproducibility and Deployment Details

### H\.1 MTP Serving Configuration

The MTP\-On point estimates are not lower on either quality dimension, so this run shows no aggregate quality degradation from enabling the adapted draft head\. Because sample\-level paired outputs and repeated decoding runs are not available, Table[16](https://arxiv.org/html/2608.15763#A6.T16)is not an equivalence test\. Table[19](https://arxiv.org/html/2608.15763#A8.T19)records the exact MTP serving configuration used for the reported measurements\.

Table 19:MTP serving configuration used for the reported quality check and deployment measurements\.
### H\.2 SGLang MTP Diagnostics

The MTP\-On SGLang traces provide scheduler\-level mechanism diagnostics, not request\-level latency\. Table[20](https://arxiv.org/html/2608.15763#A8.T20)summarizes the captured Qwen3\.6 Base window \(2026\-07\-31 09:28–09:58\) and HAT window \(2026\-07\-30 02:35–07:23\); the log timezone and platform SGLang release were not retained\. The base trace’sQwen3\_5ForCausalLMMTPstring is the platform’s internal implementation class for the checkpoint reported here as Qwen3\.6\-35B\-A3B\.

Table 20:SGLang log\-sample diagnostics for MTP\-On serving\. Triples are mean/P50/P95; Queue is P95/max\.The HAT trace’s mean 2\.03 and median 1 running request are consistent with the application’s low\-concurrency provisioning, although the P95 of 7 shows occasional bursts\. From Base to HAT, mean accept length moves from 2\.94 to 3\.12 and mean acceptance rate from 0\.735 to 0\.780; these diagnostics provide an additional mechanism\-level explanation for the MTP\-On cross\-checkpoint TPS difference beyond output length\. Acceptance measures speculative\-token verification, not semantic answer quality\. We omit SGLang input/gen throughput, cache occupancy, memory\-usage proxies, and CUDA Graph rates because the captured windows and cache states are not aligned tightly enough for causal On/Off comparison\.

### H\.3 Deployment Measurement Contract

Each benchmark run is bound to its run ID, policy checkpoint, Harness commit, sampled\-case manifest, serving route, concurrency, generation settings, timeout rules, client region, and start/end time\. H20 runs additionally record GPU count, TP/PP, dtype, MTP parameters, and known scheduler settings\. Request\-level records retain wall latency, first\-call TTFT, every model call’s latency and usage, final execution status, and warm\-up membership\. A future production\-traffic claim would additionally require a frozen service version, an explicit traffic window, request\-weighted concurrency and queue distributions, exclusion rules, and raw telemetry\.