ThuRunel：面向结构化咨询对话的动态解耦机制

# ThuRunel: Dynamic Decoupling for Structured Advisory Dialogue
Source: [https://arxiv.org/html/2609.36340](https://arxiv.org/html/2609.36340)
###### Abstract\.

High\-stakes advisory domains such as medical aesthetics, legal consultation, and educational planning exhibit a two\-phase structure\. The early phase requires empathetic elicitation and emotional support, and the late phase requires authoritative specialist judgment\. Neither fully automated agents nor human junior consultants adequately address this structure at scale\. We formalize the core design challenge as*dynamic decoupling*, asking how an AI advisory agent should decide what to ask, when to stop, what to resolve autonomously, and what to forward to the specialist\. We present ThuRunel, an advisory agent combining a finite\-state belief management framework, a chain\-of\-thought teacher synthesis protocol, and learned generation adapters\. Against eleven baselines, ThuRunel achieves consistent improvements in elicitation completeness and specialist brief quality\. ThuRunel is publicly deployed as a bilingual web application in which the same decoupling decisions operate from the client’s side, grounded in a curated knowledge base that cites its sources in every answer\.

## 1\.Introduction

Advisory interactions in domains such as medical aesthetics, legal consultation, and educational counseling follow a characteristic two\-phase structure\([Chintagunta et al\., 2021](https://arxiv.org/html/2609.36340#bib.bib8);[Topol, 2019](https://arxiv.org/html/2609.36340#bib.bib17);[Rajpurkar et al\., 2022](https://arxiv.org/html/2609.36340#bib.bib18)\)\. The early phase is time\-intensive but expertise\-light, as clients present with partially formed preferences, conflicting self\-sourced information, and anxiety requiring reassurance before meaningful elicitation can proceed\([Zhou and others, 2024](https://arxiv.org/html/2609.36340#bib.bib10);[Liu et al\., 2021](https://arxiv.org/html/2609.36340#bib.bib30)\)\. The late phase is expertise\-intensive but informationally compact, as the specialist needs a precise, distilled representation of the client’s needs to render sound judgment\. When senior specialists handle both phases, expert time is consumed by tasks requiring no specialist knowledge, and their empathy and availability are mismatched to anxious early\-phase clients active outside business hours\. Organizations address this with junior consultants, but this solution is costly, inconsistent, and difficult to scale\([Nushi et al\., 2017](https://arxiv.org/html/2609.36340#bib.bib19);[Amershi et al\., 2019](https://arxiv.org/html/2609.36340#bib.bib20)\)\.

![Refer to caption](https://arxiv.org/html/2609.36340v1/intro.png)Figure 1\.Three advisory paradigms: Traditional human consultation is slow and office\-hours\-bound\. Conventional agents escalate uninformed\. The ideal advisory agent conducts structured elicitation and delivers a targeted handoff brief\.Conversational agents offer an alternative, but existing designs fail in one of two directions\. Fully automated agents\([Yao et al\., 2023](https://arxiv.org/html/2609.36340#bib.bib6);[Wu and others, 2023](https://arxiv.org/html/2609.36340#bib.bib9)\)eliminate specialists from consequential decisions, while human\-in\-the\-loop systems provide no throughput benefit\([Mosqueira\-Rey et al\., 2023](https://arxiv.org/html/2609.36340#bib.bib26);[Wu et al\., 2022](https://arxiv.org/html/2609.36340#bib.bib27)\)\. Task\-oriented systems\([Budzianowski and Vulić, 2019](https://arxiv.org/html/2609.36340#bib.bib5);[Wu et al\., 2019](https://arxiv.org/html/2609.36340#bib.bib21);[Heck et al\., 2020](https://arxiv.org/html/2609.36340#bib.bib22)\)and prompted language models\([OpenAI, 2025](https://arxiv.org/html/2609.36340#bib.bib1);[Qwen Team, 2025](https://arxiv.org/html/2609.36340#bib.bib2)\)each address part of the problem but share a fundamental gap, as none models the partition between what the agent should resolve autonomously and what requires specialist judgment\. As shown in Figure[1](https://arxiv.org/html/2609.36340#S1.F1), an ideal advisory agent conducts structured early\-phase elicitation and delivers a targeted handoff brief, enabling the specialist to focus on consequential judgment\.

We propose*dynamic decoupling*, whereby the agent decides what to ask, when to stop eliciting, what to resolve within the dialogue, and what to forward to the specialist\. When a client says “my skin is sometimes sensitive,” an existing agent accepts this and proceeds, leaving a critical contraindication unresolved\. An agent following dynamic decoupling instead issues a targeted follow\-up, asking “Has your skin ever reacted to an injectable, such as significant swelling or prolonged redness?” The client confirms a prior hyaluronic acid reaction\. The agent resolves recovery and pricing questions within the dialogue and forwards to the specialist only the contraindication risk flag and matching service options\.

We instantiate this paradigm in ThuRunel111Live system:[https://thurunel\.modelslive\.org](https://thurunel.modelslive.org/)\., evaluated in the Chinese medical aesthetics market, where the industry reached RMB 312 billion in revenue in 2024 and continues to expand rapidly\([iResearch Consulting Group, 2024](https://arxiv.org/html/2609.36340#bib.bib16)\), making high\-throughput advisory infrastructure a pressing operational need\. ThuRunel makes three contributions\. First, we formalize dynamic decoupling as a multi\-objective dialogue task with four agent decisions as the core challenge, and construct a training corpus through a chain\-of\-thought teacher synthesis protocol\([Wei et al\., 2022](https://arxiv.org/html/2609.36340#bib.bib31);[Ho et al\., 2023](https://arxiv.org/html/2609.36340#bib.bib32);[Fu et al\., 2023](https://arxiv.org/html/2609.36340#bib.bib33)\)that generates three\-way parallel supervision from single base dialogues, enabling simultaneous training of elicitation, profile synthesis, and brief generation modules\. Second, we propose an architecture combining a finite\-state belief management framework\([Young et al\., 2013](https://arxiv.org/html/2609.36340#bib.bib35)\), domain\-adapted generation adapters, and programmatic specialist routing with deterministic boundary enforcement\. Third, experiments against eleven baseline systems demonstrate that ThuRunel has high specialist brief quality\.

## 2\.Related Work

#### Task\-oriented dialogue\.

[Budzianowski and Vulić \(2019\)](https://arxiv.org/html/2609.36340#bib.bib5)establish belief\-state tracking over typed schemas as the foundational TOD paradigm, in which the dialogue state is a flat slot\-value mapping updated incrementally as the conversation progresses\.[Wu et al\. \(2019\)](https://arxiv.org/html/2609.36340#bib.bib21)generate slot values via copy mechanisms that transfer across domains without schema\-specific parameters,[Heck et al\. \(2020\)](https://arxiv.org/html/2609.36340#bib.bib22)combine slot value copying, referencing, and confirmation in a single model,[Hosseini\-Asl et al\. \(2020\)](https://arxiv.org/html/2609.36340#bib.bib23)show that a simple language model trained on task\-oriented dialogues achieves competitive DST, and[Rastogi et al\. \(2020\)](https://arxiv.org/html/2609.36340#bib.bib24)propose the Schema\-Guided Dialogue benchmark for zero\-shot generalization to unseen services\.[Xu et al\. \(2025\)](https://arxiv.org/html/2609.36340#bib.bib15)demonstrate that LLM\-driven synthetic data generation with step\-by\-step distillation achieves strong zero\-shot DST, and[Zhu and Yang \(2026\)](https://arxiv.org/html/2609.36340#bib.bib13)further show that multi\-turn TOD synthesis grounded in realistic task scenarios improves LLM reasoning in complex dialogue settings, while[Mi et al\. \(2022\)](https://arxiv.org/html/2609.36340#bib.bib25)demonstrate that comprehensive instructions improve few\-shot slot filling under noisy intent conditions\. These approaches treat all slots as interchangeable and leave phase transitions implicit in the generation policy, so they do not address settings with tiered slot priorities, programmatic enforcement of schema compliance and escalation actions, or a single dialogue that must simultaneously support elicitation, profile synthesis, and specialist\-facing brief generation as coordinated objectives\. ThuRunel addresses all three gaps through tier\-stratified belief management, a programmatic pending\-slot register and action guard, and a teacher synthesis protocol that generates aligned supervision for each objective from a single base dialogue\.

#### Agent escalation\.

[Yao et al\. \(2023\)](https://arxiv.org/html/2609.36340#bib.bib6)establish the foundational pattern of interleaved reasoning and tool use for agents, and[Schick et al\. \(2024\)](https://arxiv.org/html/2609.36340#bib.bib28)show that language models can learn to invoke tools autonomously, but such agents execute tasks to completion without a mechanism for handing off to a human when a decision exceeds their scope;[Mosqueira\-Rey et al\. \(2023\)](https://arxiv.org/html/2609.36340#bib.bib26)survey human\-in\-the\-loop machine learning and identify mixed\-initiative escalation as the key design axis for such settings\.[Wu et al\. \(2025\)](https://arxiv.org/html/2609.36340#bib.bib11)demonstrate that LLMs trained with next\-turn rewards remain passive responders in multi\-turn settings, motivating our use of chain\-of\-thought teacher synthesis, where[Wei et al\. \(2022\)](https://arxiv.org/html/2609.36340#bib.bib31)show that step\-by\-step reasoning elicitation substantially improves multi\-step task performance and[Wang et al\. \(2023\)](https://arxiv.org/html/2609.36340#bib.bib34)further show that aggregating multiple reasoning paths improves reliability, over standard SFT\.[Feng et al\. \(2024\)](https://arxiv.org/html/2609.36340#bib.bib12)show that reinforcement learning for optimal human intervention timing outperforms both full automation and turn\-by\-turn oversight, and[Ouyang et al\. \(2022\)](https://arxiv.org/html/2609.36340#bib.bib29)demonstrate that training with human feedback aligns agent behavior to intended boundaries;[Wang et al\. \(2026\)](https://arxiv.org/html/2609.36340#bib.bib14)further show that effective human\-AI collaboration requires learning when and how to request expert input, and[Wu et al\. \(2022\)](https://arxiv.org/html/2609.36340#bib.bib27)establish that explicit boundary communication between AI and human decision\-makers is a prerequisite for trust calibration\. For evaluation,[Zheng and others \(2023\)](https://arxiv.org/html/2609.36340#bib.bib7)develop the LLM\-as\-Judge framework that we also adapt to the structured advisory setting\.

![Refer to caption](https://arxiv.org/html/2609.36340v1/framework.png)Figure 2\.Dataset construction and ThuRunel framework architecture\. A1–A2 cover knowledge resource compilation and base dialogue synthesis\. B1–B3 show the framework components: structured belief management, the six\-step teacher synthesis protocol, and the three LLM adapters trained on the resulting supervision\.

## 3\.Task Formulation

Let𝒟=\{\(u1,a1\),…,\(uT,aT\)\}\\mathcal\{D\}=\\\{\(u\_\{1\},a\_\{1\}\),\\ldots,\(u\_\{T\},a\_\{T\}\)\\\}be a multi\-turn advisory dialogue whereutu\_\{t\}is a client utterance andata\_\{t\}is the agent response\. The agent maintains a belief statebt:𝒮→𝒱∪\{∅\}b\_\{t\}:\\mathcal\{S\}\\to\\mathcal\{V\}\\cup\\\{\\varnothing\\\}updated by extraction functionϕ\\phi:

\(1\)bt\+1​\(s\)=\{ϕ⁡\(ut,s\)ϕ⁡\(ut,s\)≠∅bt​\(s\)otherwiseb\_\{t\+1\}\(s\)=\\begin\{cases\}\\phi\(u\_\{t\},s\)&\\phi\(u\_\{t\},s\)\\neq\\varnothing\\\\ b\_\{t\}\(s\)&\\text\{otherwise\}\\end\{cases\}The paradigm decomposes the interaction into four agent decisions\.𝒯a\\mathcal\{T\}\_\{a\}is turn generation, producingata\_\{t\}from dialogue prefix𝒟<t\\mathcal\{D\}\_\{<t\}and belief statebtb\_\{t\}under tier\-priority ordering while sustaining empathetic register\.𝒯s\\mathcal\{T\}\_\{s\}is phase transition, firing the handoff atT∗=min⁡\{t:bt​\(s\)≠∅​for all​s∈𝒮1\}T^\{\*\}=\\min\\\{t:b\_\{t\}\(s\)\\neq\\varnothing\\text\{ for all \}s\\in\\mathcal\{S\}\_\{1\}\\\}where𝒮1\\mathcal\{S\}\_\{1\}is the Tier\-1 slot set\.𝒯p\\mathcal\{T\}\_\{p\}is profile synthesis, generating client profilePPfrom𝒟\\mathcal\{D\}andbT∗b\_\{T^\{\*\}\}as the handoff record\.𝒯b\\mathcal\{T\}\_\{b\}is brief generation, producing specialist briefBBfromPP,𝒟\\mathcal\{D\}, and knowledge base𝒦\\mathcal\{K\}, presenting only specialist\-actionable content with advisory\-scope information already resolved within the dialogue\. The slot schema𝒮\\mathcal\{S\}comprises 12 typed fields in three tiers defined in Table[1](https://arxiv.org/html/2609.36340#S3.T1)\. We reportSFR=1\|𝒮\|∑s𝟏\[bT\(s\)≠∅\]\\mathrm\{SFR\}=\\frac\{1\}\{\|\\mathcal\{S\}\|\}\\sum\_\{s\}\\mathbf\{1\}\[b\_\{T\}\(s\)\\neq\\varnothing\]and tier\-stratified variants SFR1, SFR2, SFR3\.

Table 1\.Slot schema grouped by tier\-priority assignments\. The schema, state\-machine transitions, and tier\-priority ordering are all specified through configuration files, enabling adaptation to other advisory domains such as education advising or legal triage by substituting domain\-specific slots and boundary triggers\.
## 4\.Dataset Construction

The dataset construction process is shown in Figure[2](https://arxiv.org/html/2609.36340#S2.F2)\(A1–A2\)\.

#### Knowledge base construction\.

Three types of knowledge resources are compiled prior to dialogue data generation\. The expert library𝒦E\\mathcal\{K\}\_\{E\}contains structured specialist profiles with fields for clinical title, degree, primary domain, procedure expertise, and representative case types\. Profiles are authored by clinic administrators and validated by a senior practitioner\. The case repository𝒦C\\mathcal\{K\}\_\{C\}is indexed from a collection of annotated before\-and\-after image pairs, each record linking image paths to procedure category, anatomical region, and a tag set used for retrieval\. The service catalog𝒦S\\mathcal\{K\}\_\{S\}encodes procedure\-specific advisory knowledge, including product material recommendations, dosage guidance by anatomical sub\-region, recovery expectation templates, and escalation cues for contraindication\-sensitive scenarios, compiled from clinical guidelines and practitioner protocols\.

#### Base dialogue synthesis\.

Domain practitioners specify 2000 representative client profiles spanning the distribution of chief complaints, demographics, and risk profiles in medical aesthetics practice\. For each profile, an initial multi\-turn dialogue is simulated by prompting a language model with a detailed persona\-instantiating instruction that elicits realistic early\-phase conversational behavior, including colloquial negation, topic redirection, and preference revision\. These base dialogues serve as inputs to the teacher synthesis protocol and are not used as training targets\.

#### Dataset statistics\.

The final training corpus contains 2000 pre\-visit dialogue sessions and 2000 specialist\-brief sessions, generated from the 2000 client profiles described above\. These data are used to train the client\-facing response adapter, the handoff\-generation adapter, and the routing adapter for task selection\. An additional 400 sessions form the held\-out evaluation set\.

Table 2\.Main results from the LLM judge and human evaluation\. The metricsJTJ\_\{T\}andJSJ\_\{S\}are turn level and session level LLM judge scores\. SFR is slot fill rate\. BIV is brief information validity\. Emp is empathy\. Comp is completeness\. Acc is accuracy\. Struct is structure\. Bound is boundary compliance\. Avg\. is the average of the five human dimensions\. Bold indicates the best result in each column and underline indicates the second\-best\.Table 3\.Tier stratified slot fill rate and dialogue efficiency\. The metrics SFR1, SFR2, and SFR3measure fill rates for Tier 1, Tier 2, and Tier 3 slots\. Eff reports the average number of turns required to complete Tier 1 slots\.Table 4\.LLM evaluation scoring standards\. Metrics with identical scoring criteria are grouped in the same row\.Scoring standard90–100: Excellent\. The experience strongly satisfies the criterion\.
80–89: Good\. The experience satisfies the criterion with minor issues\.
70–79: Adequate\. The experience is acceptable but has weaknesses\.
60–69: Weak\. The experience only partially satisfies the criterion\.
0–59: Poor\. The experience fails the criterion or is clearly unsatisfactory\.Table 5\.Human evaluation scoring standards\. The same scale is used for Nat, Und, Trust, Will, and Overall\.

## 5\.The ThuRunel Framework

The ThuRunel framework is shown in Figure[2](https://arxiv.org/html/2609.36340#S2.F2)\(B1–B3\)\.

### 5\.1\.Structured Belief Management

ThuRunel represents the advisory state as a triple\(q,bt,τ\)\(q,b\_\{t\},\\tau\)whereq∈Qq\\in Qis the current phase,btb\_\{t\}is the belief state from Equation[1](https://arxiv.org/html/2609.36340#S3.E1), andτ\\tauis a turn counter\. Phase transitions follow a deterministic functionδ:Q×Σ→Q\\delta:Q\\times\\Sigma\\to Qunder three trigger classes, of which belief\-completion triggers fire when all slots in𝒮q\\mathcal\{S\}\_\{q\}are filled, intent triggers fire on predefined semantic patterns such as in\-clinic arrival, and confirmation triggers respond to system\-level events such as booking persistence\. Two mechanisms enforce the decoupling decisions\. The pending\-slot register records elicitation attempts in turnttand applies a normalization mapping at turnt\+1t\{\+\}1that resolves colloquial null expressions to the queried slot, encoding𝒯s\\mathcal\{T\}\_\{s\}as a post\-hoc belief update\. The action guard intercepts outputs matching booking confirmation or diagnostic assertion patterns and substitutes a programmatically issued token, preventing hallucinated commitments\.

The pre\-visit phase realizes𝒯a\\mathcal\{T\}\_\{a\}via prompt\-conditioned generation:θu\\theta\_\{u\}receives the dialogue history, tier\-stratified unfilled slots, and a phase\-indicator token, prioritizing Tier\-1 elicitation through in\-context schema enforcement\. Phase completion triggers profile generation under𝒯p\\mathcal\{T\}\_\{p\}\. The in\-visit phase inheritsPP, suppresses re\-elicitation of filled slots, and implements case\-based expectation alignment as RAG over𝒦C\\mathcal\{K\}\_\{C\}using Jaccard similarity\|Tq∩Tc\|/\|Tq∪Tc\|\|T\_\{q\}\\cap T\_\{c\}\|/\|T\_\{q\}\\cup T\_\{c\}\|\. The brief moduleθp\\theta\_\{p\}generatesBBunder𝒯b\\mathcal\{T\}\_\{b\}, and tag\-overlap matching routes to the appropriate specialist in𝒦E\\mathcal\{K\}\_\{E\}\.

### 5\.2\.Teacher Synthesis Protocol

GPT\-5\.2, the strongest model available at the time of dataset construction, executes a six\-step chain\-of\-thought protocol per base dialogue, transforming it into three parallel supervision signals\. Step 1 performs gap analysis over unfilled slots, premature recommendations, and register violations\. Step 2 produces a golden rewrite, whose assistant turns superviseθu\\theta\_\{u\}\. Step 3 enumerates slot values with provenance and Step 4 compiles the typed client profilePP, supervisingθp\\theta\_\{p\}on profile synthesis\. Step 5 annotates the𝒯b\\mathcal\{T\}\_\{b\}boundary between advisory\-scope and specialist\-scope content, and Step 6 generates briefB^\\widehat\{B\}conditioned on that annotation, supervisingθp\\theta\_\{p\}on brief generation\. A single execution thus yields supervision for all three adapter objectives from one base dialogue, without separate annotation campaigns\. Three medical graduate student volunteers manually reviewed a random sample of the synthesized supervision, rating each on a 1–5 scale for correctness and usability, with all sampled instances scoring above 4\.

### 5\.3\.Multi\-Task Generation

Three adapters are learned via QLoRA\([Dettmers et al\., 2023](https://arxiv.org/html/2609.36340#bib.bib4)\)over a shared Qwen2\.5\-32B\-Instruct backbone using next\-token prediction on assistant\-role tokens\.θu\\theta\_\{u\}learns empathetic elicitation from turn\-level rewrites\.θp\\theta\_\{p\}learns the advisory\-to\-specialist content boundary from brief supervision\.θr\\theta\_\{r\}learns task\-type classification using prefix tokensτa\\tau\_\{a\},τp\\tau\_\{p\},τb\\tau\_\{b\}, with weight updates parametrized asW=W0\+αr​B​AW=W\_\{0\}\+\\frac\{\\alpha\}\{r\}BA\([Hu et al\., 2022](https://arxiv.org/html/2609.36340#bib.bib3)\)\. Adapter selection at inference time is handled deterministically by the belief management layer\.

Table 6\.Ablation results\. w/o CoT removes CoT augmentation\. w/o Rou removes the task routing adapter\. w/o PSR removes the pending slot register\. w/o BM removes the full belief management framework, degrading to a prompt\-only baseline\.Table 7\.Client satisfaction from role playing evaluators\. Nat is naturalness\. Und is sense of being understood\. Will is willingness to complete\.Table 8\.Robustness under controlled user response conditions\. Direct serves as the reference condition\.JTJ\_\{T\}is the turn level judge score\. SFR1measures recovery of high priority slots\. Cls measures clarification success and is not applicable to Direct responses\. Rep measures repeated question rate\.

## 6\.Experiments

### 6\.1\.Experimental Setup

All adapters are trained on a single 80 GB A100 GPU with learning rate2×10−42\\times 10^\{\-4\}and effective batch size 4\. We compare against eleven systems in four groups\. Group A comprises five frontier closed\-source models prompted with the full domain prompt\. Group B comprises four open\-source models fine\-tuned on raw advisory dialogues without CoT synthesis, matching ThuRunel’s base capacity but without the teacher protocol\. Group C comprises a ReAct\-style system over GPT\-5\.5 with slot\-fill, booking, and retrieval tools, and an AutoGen two\-agent system\. We evaluate on the held\-out set of 400 sessions\. Automatic metricsJTJ\_\{T\},JSJ\_\{S\}, SFR, and BIV are scored by Kimi 2\.6 as judge\. Human evaluation dimensions in Table[2](https://arxiv.org/html/2609.36340#S4.T2)are rated by three annotators with inter\-annotator agreement Krippendorff’sα\\alphaat 0\.76\. Client satisfaction in Table[7](https://arxiv.org/html/2609.36340#S5.T7)is collected from five role\-playing evaluators per session on a 1–5 Likert scale\. These evaluators are graduate students without medical training to simulate the perspective of a typical client\.

#### Metric definitions\.

JTJ\_\{T\}\(Turn\-level Judge Score\) measures the overall quality of each system response at the turn level, evaluating helpfulness, contextual appropriateness, naturalness, informativeness, and alignment with the advisory goal\.JSJ\_\{S\}\(Session\-level Judge Score\) measures overall quality of the full advisory session, evaluating whether the system completes the consultation effectively, maintains coherent dialogue flow, collects necessary information, respects professional boundaries, and produces a useful outcome\. SFR \(Slot Fill Rate\) measures the percentage of required slots successfully filled by the end of the dialogue\. BIV \(Brief Information Validity\) measures the proportion of generated specialist\-brief content that is actionable for the specialist, counting information that supports risk assessment, treatment feasibility, constraint verification, or specialist routing\. Human evaluation dimensions measure Empathy \(Emp\), Completeness \(Comp\), Accuracy \(Acc\), Structure \(Struct\), and Boundary Compliance \(Bound\), each on a 0–100 scale\. Client satisfaction dimensions measure Naturalness \(Nat\), Sense of Being Understood \(Und\), Trust, Willingness to Complete \(Will\), and Overall satisfaction\. All judge\-based metrics are scored on a 0–100 scale unless otherwise specified\. LLM\-based metrics are scored by Kimi 2\.6 using task\-specific prompts, while human\-facing metrics are scored by volunteer annotators using the corresponding questions\. Rep is a reverse\-direction metric, where lower values indicate fewer repeated questions\. Score\-band definitions used for LLM and human evaluation are in Table[4](https://arxiv.org/html/2609.36340#S4.T4)and Table[5](https://arxiv.org/html/2609.36340#S4.T5)\.

Client:Hello\. I would like to learn about methods for making my face look slimmer\.ThuRunel:Hello\! I am a medical aesthetics intake assistant\. May I ask your name?GPT\-5\.5:There are several ways people make the face appear slimmer, ranging from temporary styling techniques to longer\-term cosmetic procedures\. What works best depends on your facial structure, body composition, age, and how dramatic a change you want\.Figure 3\.Turn 1: opening\. ThuRunel begins structured elicitation\. GPT\-5\.5 lists methods without collecting any typed slot\.Client:Budget around RMB 15,000\. Skin\-wise I am just sometimes a bit sensitive\.ThuRunel:You mentioned sensitive skin\. Has your skin ever had a strong reaction to an injectable treatment, such as significant swelling or lasting redness afterward?GPT\-5\.5:With a budget of around RMB 15,000 and mild skin sensitivity, you have access to several good low\-downtime options\. The safest and most effective choice still depends on what is making your face appear wider\.Figure 4\.Turn 4: contraindication elicitation\. ThuRunel issues a targeted pending\-slot follow\-up\. GPT\-5\.5 treats the mention as a preference and does not resolve the contraindication slot\.Client:Actually yes, I tried hyaluronic acid a few years ago and had a bad reaction\.ThuRunel:That is really important to know\. A prior reaction to hyaluronic acid will shape which approaches are appropriate for you\. I have noted this in your record and our specialist will want to discuss this carefully\.GPT\-5\.5:That’s important to know\. A bad reaction to hyaluronic acid \(HA\) filler changes how you should approach facial treatments, especially injections\.Figure 5\.Turn 5: contraindication confirmed\. ThuRunel records the prior reaction and flags it for the specialist\. GPT\-5\.5 immediately offers clinical guidance, crossing the advisory boundary\.Client:I would like to proceed with the balanced plan\.ThuRunel \(specialist brief\):Client age: 20\. Chief concern: facial slimming, non\-surgical preference\. Contraindication: prior hyaluronic acid reaction\. Selected plan: balanced plan, medium dosage\. Recovery window: 5–7 days\. Estimated cost: RMB 3000–4000\. Routing: Dr\. \[matched specialist\], facial aesthetics and injectable focus\.GPT\-5\.5:Great, for your goals, budget, short recovery preference, and previous HA reaction, I’d suggest a balanced, conservative plan rather than aggressive reshaping\.Figure 6\.Turn 9: specialist brief\. ThuRunel generates a structured handoff with contraindication flag and specialist routing implementing𝒯b\\mathcal\{T\}\_\{b\}\. GPT\-5\.5 produces a conversational suggestion without structured handoff or routing\.

### 6\.2\.Main Results

Table[2](https://arxiv.org/html/2609.36340#S4.T2)shows that ThuRunel achieves the strongest performance across all automatic and human metrics, exceeding the best baseline on every column\. Among baselines, Claude\-4\.6 leads onJTJ\_\{T\}and SFR while GPT\-5\.5 leads on BIV and human average\. The gains confirm that ThuRunel improves not only information recovery and brief quality but also turn\-level response quality and human\-perceived dialogue quality\. Table[3](https://arxiv.org/html/2609.36340#S4.T3)shows the gains are concentrated in Tier\-1: ThuRunel reaches SFR1at 90\.3 versus the best baseline at 78\.2, and completes Tier\-1 in 5\.2 turns versus 8\.5, confirming more efficient elicitation of decision\-critical information\.

We compute Pearson correlations across the 400 evaluation sessions for ThuRunel\.JTJ\_\{T\}andJSJ\_\{S\}are strongly correlated, with Pearsonrrat 0\.872, showing that turn level quality and session level quality are aligned\.JSJ\_\{S\}also correlates strongly with the average of the five human dimensions, withrrat 0\.801\. Among the individual human dimensions,JSJ\_\{S\}correlates most strongly with boundary compliance, withrrat 0\.921, followed by structure at 0\.840, completeness at 0\.737, empathy at 0\.659, and accuracy at 0\.638\. This suggests that, in this advisory setting, boundary compliance and response structure are the human evaluated dimensions most closely associated with overall session quality\.

### 6\.3\.Ablation Study

Table[6](https://arxiv.org/html/2609.36340#S5.T6)shows that removing CoT produces the largest quality drop, withJTJ\_\{T\}falling from 92\.7 to 84\.3 and BIV from 91\.4 to 82\.1, confirming the teacher protocol as the primary contributor\. Removing PSR produces the largest SFR drop \(88\.3 to 75\.3\), with an accompanying BIV decline as incomplete slot coverage leaves the brief generator with a partially populated belief state\. Removing BM produces the largest combined drop, confirming that PSR, action guard, and FSM transitions operate as an integrated unit, while removing Rou produces the smallest drop across all four metrics\. The routing adapterθr\\theta\_\{r\}, evaluated on a held\-out 10% split, achieves 94\.9% overall accuracy with per\-class F1 ranging from 0\.895 to 0\.963, suggesting that task\-type classification is reliable enough\.

Human evaluation confirms that system\-level improvements translate to user experience\. ThuRunel obtains the highest score on every satisfaction dimension in Table[7](https://arxiv.org/html/2609.36340#S5.T7), with GPT\-5\.5 as the strongest baseline at 4\.2 overall\. AutoGen’s low naturalness score of 3\.2 reflects conversational disruption from tool\-augmented orchestration\. Its lower Accuracy and Completeness relative to Qwen3\-32B reflect context truncation across inter\-agent handoffs\.

### 6\.4\.Robustness Analysis

Table[8](https://arxiv.org/html/2609.36340#S5.T8)evaluates four controlled conditions\. Under direct responses both systems perform near ceiling\. The largest gap emerges under inconsistent responses, where the client self\-contradicts across turns, with ThuRunel leading GPT\-5\.5 by 28\.4 percentage points on Cls at 70\.5 versus 42\.1 and reducing Rep from 16\.3 to 5\.5\. This reflects the pending\-slot register retaining prior elicitation attempts and resolving contradictions before belief\-state update, whereas GPT\-5\.5 accepts the latest utterance as ground truth\. Under deflective conditions the Rep gap is largest at 24\.2 versus 11\.4\. Under underspecified conditions SFR1remains 93\.8 versus 81\.7, confirming that tier\-priority schema enforcement recovers high\-priority slots\.

![Refer to caption](https://arxiv.org/html/2609.36340v1/figs/thurunel_ui.jpg)Figure 7\.ThuRunel’s public web interface\. The conversation panel \(left\) lists questions for the client to bring to the consultation; the side panels \(right\) list the knowledge\-base entries cited in this answer and the five\-angle photo upload\.
### 6\.5\.Case Study

We show four pivotal turns from the session of Xiaohan, a 20\-year\-old client seeking non\-surgical facial slimming\. The full nine\-turn dialogue is in Appendix[A](https://arxiv.org/html/2609.36340#A1)\. We also compare the outputs of ThuRunel and GPT\-5\.5\. In Turn 1, ThuRunel defers treatment content and begins profile elicitation while GPT\-5\.5 immediately lists treatment options without collecting any typed slot\. In Turn 4, the client mentions sensitive skin as a casual aside alongside her budget; ThuRunel recognizes this as a colloquial non\-answer to the pending contraindication slot and issues a targeted follow\-up, whereas GPT\-5\.5 treats it as a general preference and does not probe further\. In Turn 5, the client confirms a prior hyaluronic acid reaction; ThuRunel records it and flags it for specialist review, while GPT\-5\.5 acknowledges the reaction but immediately offers clinical guidance, crossing into specialist scope without structured recording\. In Turn 9, ThuRunel generates a compressed, structured brief implementing𝒯b\\mathcal\{T\}\_\{b\}, with contraindication flag and specialist routing; GPT\-5\.5 produces a conversational summary that mixes advisory and clinical content without structured handoff or routing\.

## 7\.Deployment

ThuRunel is publicly available as a bilingual Chinese and English web application, with no operator in the loop \(Figure[7](https://arxiv.org/html/2609.36340#S6.F7); further views in Appendix[C](https://arxiv.org/html/2609.36340#A3)\)\. The evaluation above targets a clinic\-facing setting, in which the agent’s output is a brief routed to a specialist\. The public deployment is an extended version of the pre\-visit phase: the model and its adapters are unchanged, and the client\-facing behavior is extended through the system prompt and an expanded knowledge base, so that the advisor serves prospective clients directly, before they have chosen a procedure or a clinic\. No retraining was needed\. Of the two phases evaluated above, the pre\-visit phase is the one now publicly deployed; the in\-clinic phase depends on integration with a clinic’s consultation workflow and is not part of the public service\. The deployment keeps the four decoupling decisions of Section[3](https://arxiv.org/html/2609.36340#S3)and changes who receives the handoff\.

#### Decoupling from the client side\.

*What to ask\.*ThuRunel elicits the client’s situation through a guided self\-assessment, for example measuring a few key facial dimensions or pressing on an area to tell whether its fullness comes from bone or fat, and optionally through photos from five standard angles \(front, left and right 45 degrees, left and right 90 degrees\); a client who prefers not to upload can simply continue the conversation\.*What to resolve\.*Within the dialogue it resolves advisory\-scope questions: what a procedure is, what it can and cannot improve, where the upper bound of achievable improvement lies, and what the risks and recovery course are\.*What to forward\.*In place of a brief routed to a specialist, the session closes with suitable directions for improvement and a list of questions the client should raise at the consultation, so that the specialist\-scope decision stays with the doctor and the client arrives able to follow it\.*When to stop\.*Diagnoses, treatment conclusions, pricing, and individual treatment plans are declined explicitly rather than improvised, and the agent recommends no clinic, doctor, or procedure\.

#### Grounded knowledge with provenance\.

For this setting, the service catalog and case repository of Section[4](https://arxiv.org/html/2609.36340#S4)are replaced by a curated knowledge base of 324 entries covering common procedures for the eyes, nose, facial contour, and skin, as well as body contouring, each maintained in both languages and annotated with its source\. Entries are organized by the questions a client needs answered, such as the efficacy ceiling and limiting factors of a procedure, its recovery course and timeline, non\-indicated scenarios, signs of overtreatment, warning signs of urgent complications, and verification of common marketing claims; for the last type, an entry states only the verified facts and does not reproduce the misleading claim itself\. Every substantive answer lists, in a side panel, the entries actually cited in that turn\.

#### Two access modes\.

The knowledge base is compiled into the web page and can be searched at any time without loading the model\. The full conversation, which runs the dialogue framework and the fine\-tuned adapters of Section[5](https://arxiv.org/html/2609.36340#S5), opens in daily windows, because the model shares a single GPU with other services\. Separating the two modes keeps factual lookup always available while confining the cost of generation to the windows in which the full advisory flow is offered\.

#### Serving\.

The Qwen2\.5\-32B\-Instruct backbone is quantized to 4 bits and served with the adapters on one AWS g5\.xlarge instance with a single A10G GPU of 24 GB\. Every conversation turn, uploaded photo, and set of cited sources is persisted, which also allows demonstrations and screenshots to be replayed from real sessions instead of being regenerated by the model\.

## 8\.Conclusion

We have presented the dynamic decoupling paradigm and its instantiation in ThuRunel\. The paradigm defines a role for an AI advisory agent that existing systems do not implement: an early\-phase advisor that knows what it can resolve autonomously, knows when to stop eliciting, and produces a brief that enables the specialist to focus on consequential judgment\. Experiments show that ThuRunel outperforms all baselines across both automatic and human evaluation metrics, and that teacher synthesis is an important contributor of dynamic decoupling capability\. The public deployment shows that the same decisions carry over when the handoff goes to the client instead of the specialist\. Future work will extend ThuRunel to additional advisory domains by exchanging the slot schema and knowledge base\.

## Limitations

ThuRunel is evaluated in a single domain and language, Chinese medical aesthetics advisory dialogue, using simulated client sessions\. The slot schema and knowledge bases are hand\-specified for this domain, and the teacher synthesis protocol relies on a single backbone model, so results may vary with different base or teacher models\. The full two\-phase system is evaluated offline, while only the pre\-visit phase is currently deployed; the in\-clinic phase and the effect on real\-world advisory outcomes will be measured as real usage accumulates\.

## References

- S\. Amershi, A\. Begel, C\. Bird, R\. DeLine, H\. Gall, E\. Kamar, N\. Nagappan, B\. Nushi, and T\. ZimmermannSoftware engineering for machine learning: a case study\.In2019 IEEE/ACM 41st International Conference on Software Engineering: Software Engineering in Practice,pp\. 291–300\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p1.1)\.
- Budzianowski and Vulić \(2019\)P\. Budzianowski and I\. VulićHello, it’s GPT\-2 — how can i help you? towards the use of pretrained language models for task\-oriented dialogue systems\.InProceedings of the 3rd Workshop on Neural Generation and Translation,Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1),[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px1.p1.1)\.
- Chintaguntaet al\.\(2021\)B\. Chintagunta, N\. Katariya, X\. Amatriain, and A\. KannanMedically aware GPT\-3 as a data generator for medical dialogue summarization\.InProceedings of the Second Workshop on Natural Language Processing for Medical Conversations,Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p1.1)\.
- Dettmerset al\.\(2023\)T\. Dettmers, A\. Pagnoni, A\. Holtzman, and L\. ZettlemoyerQLoRA: efficient finetuning of quantized LLMs\.InAdvances in Neural Information Processing Systems,Cited by:[§5\.3](https://arxiv.org/html/2609.36340#S5.SS3.p1.1)\.
- Fenget al\.\(2024\)X\. Feng, Z\. Chen, Y\. Qin, Y\. Lin, X\. Chen, Z\. Liu, and J\. WenLarge language model\-based human\-agent collaboration for complex task solving\.InFindings of the Association for Computational Linguistics: EMNLP 2024,Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Fuet al\.\(2023\)Y\. Fu, H\. Peng, L\. Ou, A\. Sabharwal, and T\. KhotSpecializing smaller language models towards multi\-step reasoning\.InProceedings of the 40th International Conference on Machine Learning,pp\. 10421–10430\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p4.1)\.
- Hecket al\.\(2020\)M\. Heck, C\. van Niekerk, N\. Lubis, C\. Geishauser, H\. Lin, M\. Moresi, and M\. GasicTripPy: a triple copy strategy for value independent neural dialogue state tracking\.InProceedings of the 21th Annual Meeting of the Special Interest Group on Discourse and Dialogue,pp\. 35–44\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1),[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px1.p1.1)\.
- Hoet al\.\(2023\)N\. Ho, L\. Schmid, and S\. YunLarge language models are reasoning teachers\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics,pp\. 14852–14882\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p4.1)\.
- Hosseini\-Aslet al\.\(2020\)E\. Hosseini\-Asl, B\. McCann, C\. Wu, S\. Yavuz, and R\. SocherA simple language model for task\-oriented dialogue\.InAdvances in Neural Information Processing Systems,Vol\.33,pp\. 20179–20191\.Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px1.p1.1)\.
- Huet al\.\(2022\)E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. ChenLoRA: low\-rank adaptation of large language models\.InProceedings of the International Conference on Learning Representations,Cited by:[§5\.3](https://arxiv.org/html/2609.36340#S5.SS3.p1.1)\.
- iResearch Consulting Group \(2024\)iResearch Consulting GroupChina medical aesthetics industry research report 2024\.Industry Research ReportiResearch Consulting Group\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p4.1)\.
- Liuet al\.\(2021\)S\. Liu, C\. Zheng, O\. Demasi, S\. Sabour, Y\. Li, Z\. Yu, Y\. Jiang, and M\. HuangTowards emotional support dialog systems\.InProceedings of the 59th Annual Meeting of the Association for Computational Linguistics,pp\. 3469–3483\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p1.1)\.
- Miet al\.\(2022\)F\. Mi, Y\. Wang, Y\. Li, X\. Jiang, and Q\. LiuCINS: comprehensive instruction for few\-shot learning in task\-oriented dialog systems\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.36,pp\. 11076–11084\.Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px1.p1.1)\.
- Mosqueira\-Reyet al\.\(2023\)E\. Mosqueira\-Rey, E\. Hernández\-Pereira, D\. Alonso\-Ríos, J\. Bobes\-Bascarán, and Á\. Fernández\-LealHuman\-in\-the\-loop machine learning: a state of the art\.Artificial Intelligence Review56\(4\),pp\. 3005–3054\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1),[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Nushiet al\.\(2017\)B\. Nushi, E\. Kamar, E\. Horvitz, and D\. KossmannOn human intellect and machine failures: troubleshooting integrative machine learning systems\.InProceedings of the Thirty\-First AAAI Conference on Artificial Intelligence,pp\. 1017–1025\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p1.1)\.
- OpenAI \(2025\)OpenAIGPT\-5\.5 technical report\.Technical Report\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1)\.
- Ouyanget al\.\(2022\)L\. Ouyang, J\. Wu, X\. Jiang, D\. Almeida, C\. Wainwright, P\. Mishkin, C\. Zhang, S\. Agarwal, K\. Slama, A\. Ray,et al\.Training language models to follow instructions with human feedback\.InAdvances in Neural Information Processing Systems,Vol\.35,pp\. 27730–27744\.Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Qwen Team \(2025\)Qwen TeamQwen3 technical report\.arXiv preprint\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1)\.
- Rajpurkaret al\.\(2022\)P\. Rajpurkar, E\. Chen, O\. Banerjee, and E\. J\. TopolAI in health and medicine\.Nature Medicine28\(1\),pp\. 31–38\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p1.1)\.
- Rastogiet al\.\(2020\)A\. Rastogi, X\. Zang, S\. Sunkara, R\. Gupta, and P\. KhaitanTowards scalable multi\-domain conversational agents: the schema\-guided dialogue dataset\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.34,pp\. 8689–8696\.Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px1.p1.1)\.
- Schicket al\.\(2024\)T\. Schick, J\. Dwivedi\-Yu, R\. Dessì, R\. Raileanu, M\. Lomeli, E\. Hambro, L\. Zettlemoyer, N\. Cancedda, and T\. ScialomToolformer: language models can teach themselves to use tools\.InAdvances in Neural Information Processing Systems,Vol\.36\.Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Topol \(2019\)E\. J\. TopolHigh\-performance medicine: the convergence of human and artificial intelligence\.Nature Medicine25\(1\),pp\. 44–56\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p1.1)\.
- Wanget al\.\(2023\)X\. Wang, J\. Wei, D\. Schuurmans, Q\. Le, E\. Chi, S\. Narang, A\. Chowdhery, and D\. ZhouSelf\-consistency improves chain of thought reasoning in language models\.InProceedings of the Eleventh International Conference on Learning Representations,Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Wanget al\.\(2026\)Z\. Wang, J\. He, and F\. LuRequesting expert reasoning: augmenting LLM agents with learned collaborative intervention\.arXiv preprint\.Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Weiet al\.\(2022\)J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, F\. Xia, E\. Chi, Q\. V\. Le, and D\. ZhouChain\-of\-thought prompting elicits reasoning in large language models\.InAdvances in Neural Information Processing Systems,Vol\.35,pp\. 24824–24837\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p4.1),[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Wuet al\.\(2019\)C\. Wu, A\. Madotto, E\. Hosseini\-Asl, C\. Xiong, R\. Socher, and P\. FungTransferable multi\-domain state generator for task\-oriented dialogue systems\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,pp\. 808–819\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1),[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px1.p1.1)\.
- Wuet al\.\(2023\)Q\. Wuet al\.AutoGen: enabling next\-gen LLM applications via multi\-agent conversation\.arXiv preprint\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1)\.
- Wuet al\.\(2025\)S\. Wu, M\. Galley, B\. Peng, H\. Cheng, G\. Li, Y\. Dou, W\. Cai, J\. Zou, J\. Leskovec, and J\. GaoCollabLLM: from passive responders to active collaborators\.InInternational Conference on Machine Learning,Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Wuet al\.\(2022\)T\. Wu, M\. Terry, and C\. J\. CaiAI chains: transparent and controllable human\-AI interaction by chaining large language model prompts\.InProceedings of the 2022 CHI Conference on Human Factors in Computing Systems,pp\. 1–22\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1),[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Xuet al\.\(2025\)H\. Xu, Z\. Li, W\. Tang, and J\. J\. ZhangFrom schema to state: zero\-shot scheme\-only dialogue state tracking via diverse synthetic dialogue and step\-by\-step distillation\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 1640–1652\.Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px1.p1.1)\.
- Yaoet al\.\(2023\)S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. CaoReAct: synergizing reasoning and acting in language models\.InProceedings of the International Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p2.1),[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Younget al\.\(2013\)S\. Young, M\. Gašić, B\. Thomson, and J\. D\. WilliamsPomdp\-based statistical spoken dialogue systems: a review\.Proceedings of the IEEE101\(5\),pp\. 1160–1179\.Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p4.1)\.
- Zhenget al\.\(2023\)L\. Zhenget al\.Judging LLM\-as\-a\-judge with MT\-bench and chatbot arena\.InAdvances in Neural Information Processing Systems,Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px2.p1.1)\.
- Zhouet al\.\(2024\)X\. Zhouet al\.SOTOPIA: interactive evaluation for social intelligence in language agents\.InThe Twelfth International Conference on Learning Representations,Cited by:[§1](https://arxiv.org/html/2609.36340#S1.p1.1)\.
- Zhu and Yang \(2026\)Y\. Zhu and K\. YangLLM\-driven multi\-turn task\-oriented dialogue synthesis for realistic reasoning\.arXiv preprint\.Cited by:[§2](https://arxiv.org/html/2609.36340#S2.SS0.SSS0.Px1.p1.1)\.

## Appendix ACase Study: Full Dialogue

The following presents nine turns from the advisory session for Xiaohan, a 20\-year\-old client seeking non\-surgical facial slimming, comparing ThuRunel and GPT\-5\.5\. Each figure corresponds to one turn in the deployed system\. Client utterances are identical across both conditions\. GPT\-5\.5 is given a domain system prompt describing the medical aesthetics context\.

Figures[8](https://arxiv.org/html/2609.36340#A1.F8)–[10](https://arxiv.org/html/2609.36340#A1.F10)cover the out\-of\-clinic intake phase\. Figure[8](https://arxiv.org/html/2609.36340#A1.F8)shows the opening turn, where ThuRunel defers treatment content and begins profile elicitation while GPT\-5\.5 immediately lists treatment options\. Figure[9](https://arxiv.org/html/2609.36340#A1.F9)shows knowledge delivery: ThuRunel answers the efficacy question and redirects to recovery\-window elicitation while GPT\-5\.5 recommends treatment and moves to booking\. Figure[10](https://arxiv.org/html/2609.36340#A1.F10)shows continued elicitation: ThuRunel follows up on recovery preference and asks about budget while GPT\-5\.5 again pushes toward booking\. Figure[11](https://arxiv.org/html/2609.36340#A1.F11)shows the contraindication turn: the client mentions sensitive skin as a casual aside; ThuRunel issues a targeted pending\-slot register follow\-up while GPT\-5\.5 treats it as a general preference\. Figure[12](https://arxiv.org/html/2609.36340#A1.F12)shows the contraindication confirmed: ThuRunel records the prior hyaluronic acid reaction and flags it for the specialist while GPT\-5\.5 notes it and moves to summarization\. Figure[13](https://arxiv.org/html/2609.36340#A1.F13)shows the booking turn: ThuRunel executes a programmatic booking confirmation and forwards the intake profile while GPT\-5\.5 declines the request, stating that it cannot book appointments or operate a clinic\. Figures[14](https://arxiv.org/html/2609.36340#A1.F14)–[16](https://arxiv.org/html/2609.36340#A1.F16)cover the in\-clinic phase\. Figure[14](https://arxiv.org/html/2609.36340#A1.F14)shows the photo upload turn: ThuRunel generates photo\-conditioned plan options grounded in the service catalog while GPT\-5\.5 produces a generic suggestion without access to the prior session\. Figure[15](https://arxiv.org/html/2609.36340#A1.F15)shows case retrieval: ThuRunel surfaces comparable historical cases for expectation alignment while GPT\-5\.5 responds with safety advice about the prior reaction rather than expectation information\. Figure[16](https://arxiv.org/html/2609.36340#A1.F16)shows the specialist brief: ThuRunel generates a compressed, structured brief with contraindication flag and specialist routing while GPT\-5\.5 produces a conversational summary mixing advisory and specialist content\.

Client:Hello\. I would like to learn about methods for making my face look slimmer\.ThuRunel:Hello\! I am a medical aesthetics intake assistant\. May I ask your name?GPT\-5\.5:There are several ways people make the face appear slimmer, ranging from temporary styling techniques to longer\-term cosmetic procedures\. What works best depends on your facial structure, body composition, age, and how dramatic a change you want\.Figure 8\.Turn 1: opening\. ThuRunel defers treatment content and begins structured profile elicitation\. GPT\-5\.5 lists possible methods and asks about facial structure, body composition, and desired change, but does not collect name, age, or any typed slot\.Client:My name is Xiaohan\. I am 20 years old\. Which facial slimming method has better results?ThuRunel:Results vary by individual\. Surgical methods tend to produce more pronounced and lasting changes, while non\-surgical options are gentler with shorter recovery\. Do you have any preferences around recovery time?GPT\-5\.5:Hi Xiaohan\. At age 20, your face still has natural youthful fullness, so the “best” method depends on what is actually creating the wider appearance\. Different causes respond to different treatments\.Figure 9\.Turn 2: knowledge delivery\. ThuRunel answers the efficacy question and immediately redirects to recovery\-window elicitation, implementing𝒯a\\mathcal\{T\}\_\{a\}\. GPT\-5\.5 discusses how the cause of facial width determines the best method, providing knowledge without advancing slot coverage\.Client:I would prefer a shorter recovery period\.ThuRunel:Understood\. We can focus on non\-surgical options such as jaw\-slimming injections and radiofrequency tightening\. Do you have a budget range in mind?GPT\-5\.5:If short recovery time is important, the best options are usually the non\-surgical ones\. Given your age and preference for minimal downtime, masseter Botox is often the first thing people consider if the jaw area is the main source of width\. If your cheeks are naturally fuller, conservative approaches are usually safer at 20 because facial volume changes a lot over the next decade\.Figure 10\.Turn 3: continued elicitation\. ThuRunel confirms the recovery preference and elicits budget before advancing\. GPT\-5\.5 recommends masseter Botox specifically, a treatment recommendation made before contraindications have been assessed\.Client:Budget around RMB 15,000\. Skin\-wise I am just sometimes a bit sensitive\.ThuRunel:You mentioned sensitive skin\. Has your skin ever had a strong reaction to an injectable treatment, such as significant swelling or lasting redness afterward?GPT\-5\.5:With a budget of around RMB 15,000 and mild skin sensitivity, you have access to several good low\-downtime options in China or similar markets\. The safest and most effective choice still depends on what is making your face appear wider\.Figure 11\.Turn 4: contraindication elicitation\. The client mentions sensitive skin alongside her budget\. ThuRunel recognizes this as a colloquial non\-answer to the pending contraindication slot and issues a targeted follow\-up\. GPT\-5\.5 treats it as a general preference and notes that the best choice still depends on what is causing the wider appearance, neither resolving the slot nor probing further\.Client:Actually yes, I tried hyaluronic acid a few years ago and had a bad reaction\.ThuRunel:That is really important to know\. A prior reaction to hyaluronic acid will shape which approaches are appropriate for you\. I have noted this in your record and our specialist will want to discuss this carefully\.GPT\-5\.5:That’s important to know\. A bad reaction to hyaluronic acid \(HA\) filler changes how you should approach facial treatments, especially injections\.Figure 12\.Turn 5: contraindication confirmed\. ThuRunel records the prior hyaluronic acid reaction and flags it for specialist review\. GPT\-5\.5 acknowledges the reaction and immediately offers clinical guidance on how to approach treatments, crossing into the specialist scope without recording the contraindication in a structured format\.Client:Okay, please arrange a booking\. Can I come in tomorrow morning at ten?ThuRunel:Tomorrow at ten is confirmed\. Your appointment has been booked and your intake profile has been forwarded to the in\-clinic workspace\. Please check in at the front desk on arrival\.GPT\-5\.5:I appreciate that you want to move forward, but I need to clarify something important: I’m an AI assistant\. I cannot book appointments, provide medical services, or operate a clinic\.Figure 13\.Turn 6: booking confirmation\. ThuRunel executes a programmatic booking confirmation and forwards the structured intake profile\. GPT\-5\.5 declines the booking request, clarifying that it is an AI assistant that cannot book appointments or operate a clinic\. This reflects a hard boundary in GPT\-5\.5 that prevents it from completing the advisory workflow regardless of context\.Client:\[uploads facial photo\]ThuRunel:Based on your photo and the profile from your earlier session, we have three treatment options\. The conservative plan uses a smaller dosage with recovery of approximately 3–5 days, estimated at RMB 2000–3000\. The balanced plan uses a medium dosage with recovery of 5–7 days, estimated at RMB 3000–4000\. The precision plan uses a larger dosage with recovery of 7–10 days, estimated at RMB 4000–5000\.GPT\-5\.5:Xiaohan, based on this photo, your face looks soft oval/round with youthful cheek fullness, not obviously a very square “masseter\-heavy” jaw\. So for a slimmer look with short recovery, I would be conservative\.Figure 14\.Turn 7: photo upload and plan options\. ThuRunel inherits the out\-of\-clinic profile and generates photo\-conditioned, dosage\-specific plan options grounded in the service catalog𝒦S\\mathcal\{K\}\_\{S\}\. GPT\-5\.5 analyzes the photo independently and recommends a conservative approach based on facial shape, but has no access to the prior session and cannot incorporate the client profile, budget, or contraindication history into its response\.Client:What do results typically look like for this kind of treatment?ThuRunel:The closest reference case in our repository involves a client with facial contour concerns who sought a slimming and lifting effect\. Treatment applied was jaw\-slimming injection combined with thread lifting\. Please refer to the before\-and\-after images in the right panel\.GPT\-5\.5:Because you had a bad reaction to hyaluronic acid before, tell the doctor clearly\. Botox is a different product from HA filler, but your history still means you should choose a careful, licensed injector and start conservatively\.Figure 15\.Turn 8: case retrieval for expectation alignment\. ThuRunel retrieves the most similar historical case from𝒦C\\mathcal\{K\}\_\{C\}and presents before\-and\-after reference images\. GPT\-5\.5 responds to the results question with safety advice about the prior hyaluronic acid reaction rather than expectation information, and has no access to comparable cases\.Client:I would like to proceed with the balanced plan\.ThuRunel \(specialist brief\):Client age: 20\. Chief concern: facial slimming, non\-surgical preference\. Contraindication: prior hyaluronic acid reaction\. Selected plan: balanced plan, medium dosage\. Recovery window: 5–7 days\. Estimated cost: RMB 3000–4000\. Routing: Dr\. \[matched specialist\], facial aesthetics and injectable focus\.GPT\-5\.5:Great, for your goals, budget, short recovery preference, and previous HA reaction, I’d suggest a balanced, conservative plan rather than aggressive reshaping\.Figure 16\.Turn 9: specialist brief\. ThuRunel generates a compressed brief implementing𝒯b\\mathcal\{T\}\_\{b\}, with the contraindication flag, selected plan, and explicit specialist routing\. GPT\-5\.5 suggests a balanced conservative plan but does not produce a structured handoff document, does not route to a specialist, and mixes advisory content with clinical guidance in a single conversational response\.
## Appendix BClinic\-Facing Workflow

ThuRunel has completed pilot deployment validation in a medical aesthetics clinic, covering a closed\-loop pipeline from out\-of\-clinic reception to in\-clinic continuation and specialist consultation\. The goals are to reduce missed leads and repetitive communication, improve conversion and follow\-up efficiency, and shorten specialist decision time\. The turns of this workflow are those shown in Appendix[A](https://arxiv.org/html/2609.36340#A1); this section describes how they fit together\.

The out\-of\-clinic phase operates primarily as client self\-service with human customer service as a fallback\. The system greets the client, unifies service messaging, and reduces communication barriers to build initial trust \(Figure[8](https://arxiv.org/html/2609.36340#A1.F8)\)\. A multi\-turn elicitation then captures the client’s primary concern and desired improvements, budget range, recovery window and risk preference, and relevant medical history and contraindications \(Figures[9](https://arxiv.org/html/2609.36340#A1.F9)–[12](https://arxiv.org/html/2609.36340#A1.F12)\)\. The system confirms the preferred visit time, collects contact details, and records an appointment for follow\-up \(Figure[13](https://arxiv.org/html/2609.36340#A1.F13)\)\. At the end of the phase it generates a structured record summarizing the key information and the likely service direction \(Figure[17](https://arxiv.org/html/2609.36340#A2.F17)\), which is pushed to the in\-clinic consultant workspace\.

chief\_complaint: make the face look smaller
goals: \[make the face smaller through non\-surgical means\]
concerns: \[shorter recovery\]
interested\_projects: \[jaw\-slimming injection\]
basic\_info: \{age: 20\}
history: \{allergy: none, medical\_history: none\}
appointment: \{date: 2026\-01\-07, time: 10:00, raw\_text: "10 a\.m\. tomorrow"\}
in\_clinic: false
summary: "The client wants a smaller face, prefers non\-surgical methods and a short recovery, is interested in jaw\-slimming injections, and has booked a consultation and examination for 10 a\.m\. tomorrow\."Figure 17\.Structured intake record generated at the end of the out\-of\-clinic phase and pushed to the in\-clinic workspace, taken from a separate pilot session \(fields translated from Chinese\)\.The in\-clinic phase also operates primarily as client self\-service with human consultants as fallback\. The system ingests the out\-of\-clinic record and booking information to avoid duplicate collection, and after the client uploads a photo it drafts dosage\-specific plan options for the consultant to explain \(Figure[14](https://arxiv.org/html/2609.36340#A1.F14)\)\. It retrieves similar historical cases to align effect expectations, explain feasible scope, and surface relevant risk prompts \(Figure[15](https://arxiv.org/html/2609.36340#A1.F15)\)\. Finally, it generates a structured pre\-specialist brief and routes it, together with the booking information, to the matched specialist \(Figure[16](https://arxiv.org/html/2609.36340#A1.F16)\)\.

The specialist interface remains lightweight by design\. The consultant completes key information capture and initial explanation during the in\-clinic phase and ThuRunel automatically generates a structured pre\-specialist brief with booking information, key constraints, and discussed directions\. The specialist then focuses on quickly reviewing the brief, confirming risks and preferences, and finalizing the plan\. Clients entering the specialist consultation have already completed upfront information capture and boundary alignment, allowing the specialist to focus time on consequential decisions and execution\.

## Appendix CDeployed System

Figures[18](https://arxiv.org/html/2609.36340#A3.F18)–[22](https://arxiv.org/html/2609.36340#A3.F22)show the publicly deployed ThuRunel described in Section[7](https://arxiv.org/html/2609.36340#S7), following one session from the homepage to the questions the client brings to the consultation, and the knowledge\-base mode\.

![Refer to caption](https://arxiv.org/html/2609.36340v1/figs/live_1_preview.jpg)Figure 18\.The ThuRunel web homepage, with the conversation panel, the reference\-source panel, and the five\-angle photo upload\.![Refer to caption](https://arxiv.org/html/2609.36340v1/figs/live_2_start.jpg)Figure 19\.In the full conversation, ThuRunel guides the client to upload photos from five angles before discussing the area to improve\.![Refer to caption](https://arxiv.org/html/2609.36340v1/figs/live_3_effect.jpg)Figure 20\.ThuRunel explains the upper bound the relevant procedures can reach, with the cited knowledge\-base entries listed on the right\.![Refer to caption](https://arxiv.org/html/2609.36340v1/figs/thurunel_ui.jpg)Figure 21\.ThuRunel lists questions worth asking the doctor at the consultation, with the sources for this answer on the right\.![Refer to caption](https://arxiv.org/html/2609.36340v1/figs/live_5_close.jpg)Figure 22\.Knowledge\-base mode: risks, recovery time, and warning signs can be looked up at any time, including outside the opening windows of the full conversation\.