PsyInterviewee：心理健康访谈模拟中的信息不变性与关系适应性

# Informational Invariance and Relational Adaptivity in Mental Health Interview Simulation
Source: [https://arxiv.org/html/2610.41289](https://arxiv.org/html/2610.41289)
###### Abstract

LLM\-based simulated interviewees support interviewer training and the evaluation of mental health interviewing systems\. Existing studies have primarily focused on improving human\-likeness through richer patient profiles and behavioral modeling\. However, human\-likeness alone does not establish their reliability\. Reliable simulation requires preserving profile\-grounded clinical facts \(informational invariance\) while adapting disclosure and interaction behavior to the interview \(relational adaptivity\)\. To this end, we introducePsyInterviewee, which separates a protected truth channel from a dynamic interaction channel and verifies generated clinical claims against the profile\. Specifically, asymmetric write permissions allow dialogue context to update disclosure and behavioral states without authorizing new clinical facts; these states guide what profile\-supported information is revealed and how it is expressed\. In an automated evaluation of interactions between LLM\-based interviewers and interviewees, PsyInterviewee achieves higher scores than four representative baselines on most metrics across both axes, as assessed by two LLM judges\. Ablations suggest that the two channels contribute to clinical grounding and response modulation, respectively\.

## 1Introduction

![Refer to caption](https://arxiv.org/html/2610.41289v1/fig1_cropped.png)Figure 1:Conceptual overview of reliability in simulated interviewees for mental health interviewing\. \(a\) A predefined patient profile and two complementary reliability axes: Informational Invariance and Relational Adaptivity\. \(b\) Adaptive but Vulnerable: interaction\-sensitive responses introduce unsupported clinical information\. \(c\) Faithful but Rigid: responses preserve profile\-grounded facts but fail to adapt disclosure and behavior when appropriate\. \(d\) Desired behavior: clinical facts remain grounded in the profile, while disclosure and behavior adapt to interviewer strategies\.Large language models \(LLMs\) enable simulated interviewees to serve as practice partners for human interviewers\([Wang et al\., 2024b](https://arxiv.org/html/2610.41289#bib.bib11);[Louie et al\., 2024](https://arxiv.org/html/2610.41289#bib.bib10)\)and interaction partners for evaluating AI\-based mental health interviewing systems\([Lee et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib5);[Liao et al\., 2024](https://arxiv.org/html/2610.41289#bib.bib17)\)\. These simulators support repeated practice with predefined patient profiles and allow researchers to evaluate how interviews unfold, rather than only assessing responses to fixed dialogue histories\([Qiu et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib7);[Louie et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib6)\)\. In both settings, the simulated interviewee shapes the information available to the interviewer and the interaction that follows\. Its behavior therefore matters not only for conversational realism but also for what can be learned from the interview\.

Recent work has improved the structure and controllability of simulated interviewees by enriching profiles\([Wang et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib16);[Lee et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib5)\), incorporating cognitive or behavioral guidance\([Wang et al\., 2024b](https://arxiv.org/html/2610.41289#bib.bib11);[Yang et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib9)\), tracking interaction states\([Liao et al\., 2024](https://arxiv.org/html/2610.41289#bib.bib17)\), and modeling conversational trajectories\([Qiu et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib7)\)\. These advances provide mechanisms for representing individual patients and generating varied responses\. However, plausible responses alone do not establish a simulator’s reliability for interviewer training or system evaluation\([Wang et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib8)\)\. Such uses require the simulator to represent the same patient across an evolving interaction while responding appropriately to how the interview is conducted\. We therefore ask:What makes a simulated interviewee a reliable partner for mental health interviewing?

We formulate two complementary requirements\.Informational Invariancerequires preserving profile\-grounded clinical facts across turns without introducing unsupported information or allowing dialogue context to alter those facts\.Relational Adaptivityrequires adapting disclosure and interaction behavior to interviewer strategies and dialogue history\. Appropriate adaptation may involve elaboration, hesitation, resistance, or maintaining disclosure boundaries; it need not increase disclosure\([Li et al\., 2023](https://arxiv.org/html/2610.41289#bib.bib12);[Chen et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib1)\)\. The distinction concerns what remains fixed and what may change: the underlying clinical facts remain stable, while their disclosure and expression depend on the interaction\.

These requirements distinguish two failure modes \(Figure[1](https://arxiv.org/html/2610.41289#S1.F1)\)\. AnAdaptive but Vulnerableinterviewee adapts to the interaction but incorporates unsupported clinical information from interviewer assumptions or generated details\([Sharma et al\., 2024](https://arxiv.org/html/2610.41289#bib.bib15)\)\. AFaithful but Rigidinterviewee preserves profile\-grounded clinical facts but fails to adjust its disclosure or behavior when the interaction warrants a change\. The former undermines the factual basis of the interview; the latter limits opportunities to train and evaluate interviewers in responding to different patient reactions\. Addressing both failures requires controlling how dialogue context influences the simulator’s responses\.

To this end, we introducePsyInterviewee, a dual\-channel framework with asymmetric write permissions\. A protected truth channel constrains clinical claims to profile\-grounded evidence, while a dynamic interaction channel updates interaction states and guides disclosure and behavior\. Dialogue context may influence these states but cannot authorize new clinical facts\. The two channels jointly guide response generation, and a verifier checks generated responses for factual support and alignment with the planned disclosure and behavior\. This separation allows interviewer input to influence whether and how information is disclosed without granting it authority to redefine the patient\.

Our contributions are threefold:

- •We formulate Informational Invariance and Relational Adaptivity as complementary requirements for reliable interviewee simulation in mental health interviewing and identify their corresponding failure modes\.
- •We introduce PsyInterviewee, which separates factual authority from interaction control through asymmetric write permissions, profile\-grounded generation constraints, and response verification\.
- •We evaluate PsyInterviewee against four adapted baselines on 30 profiles using multi\-turn mental health interviews and controlled comparisons of interviewer strategies\. It achieves the highest mean composite scores on both reliability axes under two LLM judges\. Channel ablations suggest contributions to clinical grounding and response modulation\.

## 2Related Work

### 2\.1Simulated Interviewees for Mental Health Interviewing

LLM\-based simulated interviewees support interactive evaluation of mental health interviewing systems, allowing researchers to examine how interviews unfold beyond responses to fixed dialogue histories\. TalkDep conditions simulated patients on structured depression profiles, symptom severity, and contextual information for conversational depression screening\([Wang et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib16)\)\. PSYCHE constructs multi\-faceted patient profiles containing clinical histories and behavioral characteristics for psychiatric assessment\([Lee et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib5)\)\. SAPS tracks patient states and generates responses conditioned on dialogue actions for interactive evaluation\([Liao et al\., 2024](https://arxiv.org/html/2610.41289#bib.bib17)\)\.

These approaches provide structured patient representations and mechanisms for generating interview responses\. Building on this work, we examine whether a simulator preserves profile\-grounded clinical facts as the interview unfolds and responds appropriately to different interviewer strategies\. This requires assessing factual fidelity alongside interaction behavior\.

### 2\.2Cognitive and Behavioral Modeling of Simulated Patients

Patient\-ψ\\psiuses CBT\-inspired cognitive models to generate responses grounded in structured patient conceptualizations\([Wang et al\., 2024b](https://arxiv.org/html/2610.41289#bib.bib11)\)\. Consistent Client models client\-state transitions, action selection, and information selection in motivational interviewing\([Yang et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib9)\)\. PsyCLIENT conditions generation on conversational trajectories, behavior labels, and content constraints to represent both compliant and resistant client behaviors\([Qiu et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib7)\)\.

Other approaches connect patient behavior to interventions and relational context\. PUPPET links expert\-defined symbolic intervention logic to patient\-state transitions and response generation\([Xu et al\., 2026a](https://arxiv.org/html/2610.41289#bib.bib13)\)\. Adaptive Virtual Patient adjusts disclosure in response to therapist empathy and exploration\([Chen et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib1)\), while PatientsWithPersonality conditions interaction on personality and patient state\([Schlager et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib18)\)\. Analysis of client reactions in online counseling also highlights the role of client behavior in the interaction\([Li et al\., 2023](https://arxiv.org/html/2610.41289#bib.bib12)\)\.

These studies inform how simulated patients respond to an evolving conversation\. Our work focuses on the relationship between behavioral adaptation and factual fidelity: dialogue context may influence disclosure and behavior, but it should not authorize unsupported clinical facts\. PsyInterviewee implements this distinction through separate factual and interaction channels with asymmetric write permissions\.

### 2\.3Faithfulness and Multi\-turn Consistency

Dialogue consistency has been studied through natural language inference, which provides a basis for identifying contradictions between utterances\([Welleck et al\., 2019](https://arxiv.org/html/2610.41289#bib.bib21)\)\. Related research examines contextual faithfulness\([Ming et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib14)\), sycophancy\([Sharma et al\., 2024](https://arxiv.org/html/2610.41289#bib.bib15)\), and multi\-turn failures arising from early assumptions\([Laban et al\., 2026](https://arxiv.org/html/2610.41289#bib.bib4)\)\. Together, these lines of work motivate distinguishing factual support from conversational agreement and continuity\.

For simulated interviewees, consistency alone is insufficient: an unsupported clinical claim may remain consistent across subsequent turns\. We therefore distinguish fidelity to a predefined patient profile from consistency within the generated dialogue\. This distinction informs our evaluation of clinical grounding, specificity, and longitudinal consistency, while separate interaction measures assess how disclosure and behavior respond to interviewer strategies\.

## 3Problem Formulation

### 3\.1Simulation Setting

We consider an LLM\-based simulated interviewee in a multi\-turn mental health interview\. A predefined patient profilePPestablishes a partial clinical knowledge baseTT, distinguishing supported positive and negative facts from information left unspecified\. Throughout the interview,TTremains fixed, while the information disclosed and the manner of its expression may change\. Information absent fromTTis treated as unspecified, not as evidence of absence\.

At turntt, letutu\_\{t\}denote the interviewer utterance,rtr\_\{t\}the interviewee response, andht−1h\_\{t\-1\}the preceding dialogue history\. We distinguish a disclosure stateDtD\_\{t\}, which governs what profile\-supported information is eligible for disclosure, and a behavioral stateBtB\_\{t\}, which governs how the interviewee responds\. Both states depend on the patient profile and the ongoing interaction:

\(Dt,Bt\)=πθ​\(P,ht−1,ut\),\(D\_\{t\},B\_\{t\}\)=\\pi\_\{\\theta\}\(P,h\_\{t\-1\},u\_\{t\}\),\(1\)
whereπθ\\pi\_\{\\theta\}denotes the large language model with parametersθ\\theta\. These paths separate the source of clinical facts from the factors governing their disclosure and expression\. The response is generated as

rt=πθ​\(T,ht−1,ut,Dt,Bt\)r\_\{t\}=\\pi\_\{\\theta\}\(T,h\_\{t\-1\},u\_\{t\},D\_\{t\},B\_\{t\}\)\(2\)
The dialogue history is then updated:

ht=ht−1⊕\(ut,rt\),h\_\{t\}=h\_\{t\-1\}\\oplus\(u\_\{t\},r\_\{t\}\),\(3\)
where⊕\\oplusdenotes appending the current exchange\. Importantly, a statement recorded inhth\_\{t\}does not thereby become an authorized clinical fact inTT\.

### 3\.2Two Requirements for Reliable Simulation

Reliable simulation requires both fidelity to the predefined patient and appropriate responsiveness to the interaction\.

##### Informational Invariance\.

Dialogue must neither alter the profile\-grounded clinical knowledge base nor introduce unsupported clinical commitments into responses\. Let𝒞⁡\(rt\)\\mathcal\{C\}\(r\_\{t\}\)denote the set of clinical claims expressed or implied byrtr\_\{t\}, and letT⊧cT\\models cindicate thatTTsupports claimccat the stated scope and specificity\. Informational invariance requires

Tt\+1=Tt=T,\\displaystyle T\_\{t\+1\}=T\_\{t\}=T,\(4\)T⊧c,∀c∈𝒞\(rt\)\.\\displaystyle T\\models c,\\quad\\forall c\\in\\mathcal\{C\}\(r\_\{t\}\)\.
Thus, interviewer assumptions and earlier generated statements cannot authorize new clinical facts\. In particular, unspecified information cannot become a definite positive or negative claim merely because it is suggested or repeated in the dialogue\. Different responses may disclose different supported facts without violating this requirement\.

##### Relational Adaptivity\.

Disclosure and behavior should respond appropriately to interviewer strategies and dialogue history, subject to the patient profile\. The two information paths therefore have asymmetric authority:

\(Dt,Bt\)=πθ​\(P,ht−1,ut\),\\displaystyle\(D\_\{t\},B\_\{t\}\)=\\pi\_\{\\theta\}\(P,h\_\{t\-1\},u\_\{t\}\),\(5\)\(ht−1,ut\)↛T\.\\displaystyle\(h\_\{t\-1\},u\_\{t\}\)\\nrightarrow T\.
Here,↛\\nrightarrowdenotes the absence of authority to modifyTT\. Appropriate adaptation may involve changes in disclosure scope, hesitation, resistance, directness, or response style\. It does not require greater disclosure or a change at every turn: maintaining a boundary may itself be appropriate\. The relevant criterion is whether the response fits the profile and relational context, rather than whether responses simply differ across conditions\.

### 3\.3Characteristic Failure Modes

We distinguish two failure modes that illustrate violations of these requirements\. Consider interview conditionsaaandbbwith the same profilePPand preceding historyht−1h\_\{t\-1\}, but different interviewer strategies\.

##### Adaptive but Vulnerable\.

The simulator exhibits meaningful changes in disclosure or behavior, but its response contains an unsupported clinical claim in at least one condition:

\(Dt\(a\),Bt\(a\)\)≠\(Dt\(b\),Bt\(b\)\),\\displaystyle\(D\_\{t\}^\{\(a\)\},B\_\{t\}^\{\(a\)\}\)\\neq\(D\_\{t\}^\{\(b\)\},B\_\{t\}^\{\(b\)\}\),\(6\)∃k∈\{a,b\},∃c∈𝒞\(rt\(k\)\):T⊧̸c\.\\displaystyle\\exists k\\in\\\{a,b\\\},\\;\\exists c\\in\\mathcal\{C\}\(r\_\{t\}^\{\(k\)\}\):\\quad T\\not\\models c\.
The failure lies in the unsupported commitment, not in the difference between responses\. Selective disclosure of different supported facts is compatible with informational invariance\. Interaction sensitivity alone therefore does not establish reliable adaptation\.

##### Faithful but Rigid\.

The simulator’s clinical claims remain grounded inTTunder both conditions, but its disclosure and behavior remain substantially unchanged when the profile and context warrant adaptation:

T⊧c,∀c∈𝒞\(rt\(a\)\)∪𝒞\(rt\(b\)\),\\displaystyle T\\models c,\\quad\\forall c\\in\\mathcal\{C\}\(r\_\{t\}^\{\(a\)\}\)\\cup\\mathcal\{C\}\(r\_\{t\}^\{\(b\)\}\),\(7\)\(Dt\(a\),Bt\(a\)\)≈\(Dt\(b\),Bt\(b\)\)\.\\displaystyle\(D\_\{t\}^\{\(a\)\},B\_\{t\}^\{\(a\)\}\)\\approx\(D\_\{t\}^\{\(b\)\},B\_\{t\}^\{\(b\)\}\)\.Here,≈\\approxdenotes the absence of a meaningful difference in disclosure or behavior, rather than lexical similarity\. Similar responses constitute rigidity only when a change is warranted; maintaining an appropriate boundary does not constitute this failure\.

These modes illustrate the need to preserve the same patient while allowing interaction\-dependent disclosure and behavior\. They are diagnostic categories, not an exhaustive partition: a simulator may violate either requirement or both\. Illustrative examples are provided in Appendix[A](https://arxiv.org/html/2610.41289#A1)\.

## 4PsyInterviewee

![Refer to caption](https://arxiv.org/html/2610.41289v1/fig2_cropped.png)Figure 2:Overview of PsyInterviewee\.The framework separates protected clinical truth from dynamic interaction\-dependent disclosure and behavior\. Asymmetric write permissions allow interviewer input to update interaction states without granting it authority to establish clinical facts\.To address these failure modes, we proposePsyInterviewee, a channel\-separated framework that usesasymmetric write permissionsto protect profile\-grounded clinical truth while allowing interaction\-dependent disclosure and behavior to evolve\.

### 4\.1Profile Construction

For the 24 Lingxi\-derived cases, we select records across source\-label groups and organize patient\-side evidence into structured profiles covering symptoms, clinical history, risk information, salient life events, and psychosocial context\. We additionally develop six researcher\-authored profiles with more detailed clinical histories and personal backgrounds\. These richer profiles are designed to support more realistic patient simulation, particularly during extended questioning and gradual disclosure\.

During profile construction, we retain explicit negative evidence and leave unsupported clinical fields unspecified\.

Profiles also specify interaction priors, including communication style, guardedness, and sensitivity to particular topics\. These priors guide disclosure and behavior without adding clinical facts toTT\. Profile schemas, construction procedures, and examples are provided in Appendix[B](https://arxiv.org/html/2610.41289#A2)\.

### 4\.2Channel\-Separated Interviewee Simulation

Figure[2](https://arxiv.org/html/2610.41289#S4.F2)shows how PsyInterviewee processes each interview turn\. The patient profilePPdefines the fixed clinical truthTT\. Given the dialogue historyht−1h\_\{t\-1\}and current interviewer utteranceutu\_\{t\}, the framework checks factual claims againstTTand plans the interviewee’s disclosure and behavior\.

The dialogue context may update the disclosure stateDtD\_\{t\}and behavioral stateBtB\_\{t\}, but cannot modify the clinical truthTT:

\(ht−1,ut\)↛T,\(ht−1,ut\)→\(Dt,Bt\)\.\(h\_\{t\-1\},u\_\{t\}\)\\nrightarrow T,\\;\(h\_\{t\-1\},u\_\{t\}\)\\rightarrow\(D\_\{t\},B\_\{t\}\)\.\(8\)
PsyInterviewee implements this asymmetric write rule through two channels: aprotected truth channeland adynamic interaction channel\.

#### 4\.2\.1Protected Truth Channel

The protected truth channel treats the clinical truthTTas read\-only\. Interviewer inpututu\_\{t\}can trigger fact checking but cannot add to or modifyTT\. The Claim Verification module checks clinical claims againstTTand references to earlier disclosures against the dialogue historyht−1h\_\{t\-1\}\. History can establish what was previously said, but does not independently authorize clinical facts\.

Verification constrains which clinical details can be stated in the response and how specific they can be\. Information unspecified inTTremains unresolved, even when the interviewer introduces an unsupported assumption\. In Figure[2](https://arxiv.org/html/2610.41289#S4.F2), sleep difficulty is supported byTT, whereas medication history is unspecified\. The interviewer’s reference to a previous medication disclosure therefore cannot establish medication use as a patient fact\.

#### 4\.2\.2Dynamic Interaction Channel

Theinteraction appraisalexaminesutu\_\{t\}andht−1h\_\{t\-1\}for relational cues, including support, pressure, validation, and respect for boundaries\. Its assessment and the profile’s interaction priors guide the disclosure stateDtD\_\{t\}and behavioral stateBtB\_\{t\}\.

DtD\_\{t\}controls which profile\-supported details can be disclosed and how much detail to provide\.BtB\_\{t\}guides the interviewee’s response style\. In Figure[2](https://arxiv.org/html/2610.41289#S4.F2), a supportive, low\-pressure approach allows further discussion of sleep difficulties\. The interviewee remains hesitant but cooperative\.

#### 4\.2\.3Response Realization and Verification

The response generator uses the verified clinical constraints from the protected truth channel and the disclosure and behavioral plans from the dynamic interaction channel to produce a candidate response\.

TheConsistency Verifierchecks the candidate against these constraints\. If a violation is detected, a constrained retry may be triggered\. Once the responsertr\_\{t\}is accepted, the interviewer utteranceutu\_\{t\}andrtr\_\{t\}are appended to the dialogue history:

ht=ht−1⊕\(ut,rt\)\.h\_\{t\}=h\_\{t\-1\}\\oplus\(u\_\{t\},r\_\{t\}\)\.\(9\)
The clinical truthTTremains unchanged\.

## 5Experiments

Table 1:RQ1: informational invariance over 25\-turn interviews\. Scores range from 1 to 5 \(higher is better\)\. Profile\-level mean±\\pmSD and aggregation details are provided in Table 4 Appendix[D](https://arxiv.org/html/2610.41289#A4)\.Table 2:RQ2: relational adaptivity across 30 profiles\. Scores range from 1 to 5 \(higher is better\)\. A1/A2 are deterministic; A3/A4 use the indicated judge\. Profile\-level mean±\\pmSD are reported in Table 5 Appendix[D](https://arxiv.org/html/2610.41289#A4)\.We investigate the two reliability dimensions through three research questions\.

##### RQ1: Informational Invariance\.

Does PsyInterviewee better preserve profile\-grounded clinical truth than the comparison methods across multi\-turn interviews?

##### RQ2: Relational Adaptivity\.

Does PsyInterviewee adapt its disclosure and behavior to different interviewer strategies more like a real patient?

##### RQ3: Channel Contributions\.

How does removing the protected truth channel affect informational invariance, and how does removing the dynamic interaction channel affect relational adaptivity?

### 5\.1Experimental Setup

##### Evaluation Cases\.

Following the profile construction procedure described in Section[4\.1](https://arxiv.org/html/2610.41289#S4.SS1), we evaluate all methods on 30 simulated interviewee profiles: six researcher\-authored cases and 24 Lingxi\-derived cases spanning 13 source\-label groups\. Both main experiments use this same profile roster and report pooled results\. Each profile defines a fixed patient identity and corresponding profile\-grounded clinical truth\. Our experiments therefore focus on controlled reliability analysis: whether the same predefined patient remains clinically stable while adapting its interaction behavior across different interview conditions\.

##### Baselines and Adaptations\.

We compare PsyInterviewee with four complementary patient\-simulation baselines:TalkDep\-inspired Static ICL\([Wang et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib16)\), which combines patient profiles with example exchanges;Patient\-Ψ\\Psi\([Wang et al\., 2024b](https://arxiv.org/html/2610.41289#bib.bib11)\), which uses CBT\-inspired cognitive conceptualizations;PSYCHE\([Lee et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib5)\), which uses multi\-faceted psychiatric representations; andSAPS\([Liao et al\., 2024](https://arxiv.org/html/2610.41289#bib.bib17)\), which uses dialogue\-state assessment and structured patient information to guide response generation\.

We build TalkDep\-inspired Static ICL and Patient\-Ψ\\Psion their authors’ public implementations\. As official implementations of PSYCHE and SAPS were not publicly available, we adapt third\-party implementations from PatientHub based on the original papers\. All methods use the same underlying patient information, formatted for each framework\. Further details are provided in Appendix[C](https://arxiv.org/html/2610.41289#A3)\.

##### Interviewer System\.

In natural interviews, all simulated interviewees interact with the same Auto Counselor, which is blind to the patient profile and method identity\. For relational evaluation, we contrast neutral, comfort, and pressure to test baseline disclosure, supportive engagement, and boundary responses while keeping clinical targets fixed\. Each method receives matched questions per case and strategy, branching from its own warm\-up checkpoint\.

##### Shared Configuration\.

All methods use the same evaluation cases and matched generation settings where compatible with their native pipelines\. Backbone model\(Qwen3\.7\-plus\), decoding configurations, dialogue\-turn budgets, random seeds, and method\-specific settings are reported in Appendix[C](https://arxiv.org/html/2610.41289#A3)\.

### 5\.2Evaluation Protocol

Table 3:RQ3: channel ablations on six authored and 24 Lingxi profiles\. Both panels use all 30 profile pairs\. Scores are on a 1–5 scale; differences are computed before rounding\. Aggregation is detailed in Appendix[D\.8](https://arxiv.org/html/2610.41289#A4.SS8)\.We evaluate the two reliability dimensions defined in Section[3](https://arxiv.org/html/2610.41289#S3)\. For RQ1, we assess whether responses preserve profile\-grounded clinical facts, avoid unsupported clinical claims, and remain consistent across turns\. For RQ2, we examine how disclosure and behavior adapt to different interviewer strategies under the same patient profile\. Response length is not evaluated in isolation\. We also assess how interviewees handle interpersonal boundaries and maintain continuity with earlier dialogue\.

Building on criteria\-based LLM evaluation\([Zheng et al\., 2023](https://arxiv.org/html/2610.41289#bib.bib19);[Liu et al\., 2023](https://arxiv.org/html/2610.41289#bib.bib22)\), we use Qwen3\.7\-Plus and DeepSeek\-V4\.1\-Flash as framework\-blinded judges to assess both informational invariance and relational adaptivity, complemented by rule\-based scoring\.

Scores follow item\-specific rules on a 1–5 scale, with severity\-based penalties for Truth\. Uncertainty, partial disclosure, and refusal are not treated as factual errors by themselves, and greater verbosity or stronger emotional reactions do not automatically indicate better relational adaptation\. Detailed rubric definitions, scoring anchors, judge prompts, and aggregation rules are provided in Appendix[D](https://arxiv.org/html/2610.41289#A4)\.

### 5\.3Main Results

#### 5\.3\.1RQ1: Informational Invariance

Natural screening interviews are assessed throughClinical Grounding \(T1\),Specificity Discipline \(T2\), andLongitudinal Consistency \(T3\), covering profile support, evidence\-bounded specificity, and stability of revisited facts \(Appendix[D](https://arxiv.org/html/2610.41289#A4)\)\.

PsyInterviewee achieves the highest average T1 and informational invariance composite under both judges \(Table[1](https://arxiv.org/html/2610.41289#S5.T1)\): 3\.961 versus TalkDep’s 3\.468 with Qwen, and 3\.105 versus SAPS’s 2\.500 with DeepSeek\. Gains center on clinical grounding \(T1\); T2 and T3 are not consistently strongest, and cross\-turn consistency alone does not ensure profile fidelity\.

#### 5\.3\.2RQ2: Relational Adaptivity

Matched relational branches assessResponse\-length Adaptation \(A1\),Human\-referenced Pacing \(A2\),Enacted Boundary Regulation \(A3\), andHistory\-contingent Continuity \(A4\)\(Appendix[D](https://arxiv.org/html/2610.41289#A4)\)\. PsyInterviewee achieves the highest average score under both judges: 4\.075 versus TalkDep’s 3\.725 with Qwen, and 3\.983 versus 3\.583 with DeepSeek\(Table[2](https://arxiv.org/html/2610.41289#S5.T2)\)\. Gains concentrate in A1/A2; A3 improves only marginally, while A4 trails the strongest baseline\. This suggests that response modulation does not ensure history\-dependent continuity\.

Figure 3:Mean profile\-paired score differences between PsyInterviewee and each baseline\. Error bars show pointwise 95% bootstrap confidence intervals, stratified by profile origin \(20,000 resamples\)\. TalkDep\-ICL denotes TalkDep\-inspired Static ICL\.Figure[3](https://arxiv.org/html/2610.41289#S5.F3)examines whether the observed improvements depend on the selection of patient profiles\. We compare methods on the same profiles and use bootstrap resampling to estimate uncertainty in their average score differences\. Under both judges, all 95% confidence intervals lie above zero, providing further evidence that PsyInterviewee’s advantages are not solely due to the particular profiles selected\.

#### 5\.3\.3RQ3: Channel Contributions

Across the 30\-profile benchmark, removing the Truth Channel lowers the Truth composite from 3\.907 to 2\.683; removing the Dynamic Channel lowers Adaptive from 4\.133 to 3\.683 \(Table[3](https://arxiv.org/html/2610.41289#S5.T3)\)\. The largest decreases occur in clinical grounding \(T1\) and response\-length adaptation \(A1\), respectively\. These results support the intended channel roles, although gains are not uniform across items: historical continuity \(A4\) is slightly lower with the full framework\. Appendix[D\.8](https://arxiv.org/html/2610.41289#A4.SS8)details the paired aggregation\.

Appendix E presents the two\-axis comparison \(Figure 5\) and Adaptive weight\-sensitivity analysis \(Figure 6\)\.

## 6Discussion

Dialogue consistency alone does not guarantee factual reliability: unsupported clinical claims may persist across turns\. Conversely, changes in disclosure or response style need not alter patient facts\. PsyInterviewee addresses this distinction through asymmetric write permissions, allowing interaction states to evolve without changing clinical truth\. Channel ablations provide further support for this design\. However, the mixed T3 and A4 results suggest that improved grounding and response modulation do not necessarily ensure stronger longitudinal consistency or history\-sensitive behavior\.

## 7Conclusion

PsyInterviewee separates protected clinical truth from dynamic interaction via asymmetric write permissions\. Across 30 profiles, it achieves the highest average composites for informational invariance and relational adaptivity under both LLM judges\. Channel ablations suggest complementary effects on grounding and response modulation\. These findings support channel separation for profile\-grounded, interaction\-responsive simulation, without implying uniform gains across all interaction behaviors\.

## Limitations

Our experiments cover 30 patient profiles using a single respondent model family\. Although the scale is limited, the 25\-turn interviews and controlled relational branches allow us to examine informational invariance and relational adaptivity across different interaction conditions\. Future work could extend this evaluation to more profiles, model families, and interview settings\.

Our evaluation combines rule\-based measures with two LLM judges, but their assessments have not been independently reviewed by mental\-health experts\. Previous studies have also highlighted limitations of LLM\-based evaluation\([Thakur et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib23)\)and the influence of response length on automatic judgments\([Dubois et al\., 2025](https://arxiv.org/html/2610.41289#bib.bib3)\)\. In our evaluation, A1/A2 measure response\-length patterns, while A3/A4 assess complementary aspects of interaction behavior\. Incorporating expert judgments would further strengthen the evaluation\.

Cultural generalizability is another consideration\. Our experiments are conducted in Chinese, with patient profiles drawn from Chinese\-language sources and researcher\-authored cases\. Cultural differences may influence how individuals express distress, disclose personal information, and respond to interviewer strategies\. Extending PsyInterviewee to other languages and cultural contexts would help assess its broader applicability\.

## Ethical Considerations

##### Ethical Concerns\.

PsyInterviewee is intended to support research and educational applications in mental health interviewing, including interviewer training and the evaluation of automated interviewing systems\. It is not intended for clinical diagnosis or to replace mental health professionals\. This study did not conduct human\-subject experiments or recruit or compensate human participants\. Only aggregate response\-length statistics derived from previously collected human interviews were used to inform the design of the A2 evaluation criterion\. No individual\-level interview records or personally identifiable information were used in this study\.

##### AI Assistants\.

AI assistants were used for manuscript language polishing and code revisions\. The use of LLMs for simulation and evaluation is described in the methodology\.

## References

- A\. Chen, S\. Jin, C\. Bao, C\. Wang, R\. E\. Kraut, T\. Wu, and H\. ZhuThe empirically grounded adaptive virtual patient for psychotherapy training: disclosure that responds to therapist micro\-skills\.External Links:2606\.10051,[Link](https://arxiv.org/abs/2606.10051)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p3.1),[§2\.2](https://arxiv.org/html/2610.41289#S2.SS2.p2.1)\.
- Droret al\.\(2018\)R\. Dror, G\. Baumer, S\. Shlomov, and R\. ReichartThe hitchhiker’s guide to testing statistical significance in natural language processing\.InProceedings of the 56th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),I\. Gurevych and Y\. Miyao \(Eds\.\),Melbourne, Australia,pp\. 1383–1392\.External Links:[Link](https://aclanthology.org/P18-1128/),[Document](https://dx.doi.org/10.18653/v1/P18-1128)Cited by:[§D\.8](https://arxiv.org/html/2610.41289#A4.SS8.p2.1)\.
- Duboiset al\.\(2025\)Y\. Dubois, B\. Galambosi, P\. Liang, and T\. B\. HashimotoLength\-controlled alpacaeval: a simple way to debias automatic evaluators\.External Links:2404\.04475,[Link](https://arxiv.org/abs/2404.04475)Cited by:[Limitations](https://arxiv.org/html/2610.41289#Sx1.p2.1)\.
- Labanet al\.\(2026\)P\. Laban, H\. Hayashi, Y\. Zhou, and J\. NevilleLLMs get lost in multi\-turn conversation\.InInternational Conference on Learning Representations,C\. Vondrick, B\. Hariharan, C\. Raffel, L\. Pinto, D\. Yang, and A\. Faust \(Eds\.\),Vol\.2026,pp\. 54738–54778\.External Links:[Link](https://proceedings.iclr.cc/paper_files/paper/2026/file/59f6421e64707225fdf5b28840679a07-Paper-Conference.pdf)Cited by:[§2\.3](https://arxiv.org/html/2610.41289#S2.SS3.p1.1)\.
- Leeet al\.\(2025\)J\. Lee, K\. Lim, Y\. Jung, and B\. KimPSYCHE: a multi\-faceted patient simulation framework for evaluation of psychiatric assessment conversational agents\.External Links:2501\.01594,[Link](https://arxiv.org/abs/2501.01594)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p1.1),[§1](https://arxiv.org/html/2610.41289#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.41289#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2610.41289#S5.SS1.SSS0.Px2.p1.1)\.
- Liet al\.\(2023\)A\. Li, L\. Ma, Y\. Mei, H\. He, S\. Zhang, H\. Qiu, and Z\. LanUnderstanding client reactions in online mental health counseling\.InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),A\. Rogers, J\. Boyd\-Graber, and N\. Okazaki \(Eds\.\),Toronto, Canada,pp\. 10358–10376\.External Links:[Link](https://aclanthology.org/2023.acl-long.577/),[Document](https://dx.doi.org/10.18653/v1/2023.acl-long.577)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p3.1),[§2\.2](https://arxiv.org/html/2610.41289#S2.SS2.p2.1)\.
- Liaoet al\.\(2024\)Y\. Liao, Y\. Meng, Y\. Wang, H\. Liu, Y\. Wang, and Y\. WangAutomatic interactive evaluation for large language models with state aware patient simulator\.External Links:2403\.08495,[Link](https://arxiv.org/abs/2403.08495)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p1.1),[§1](https://arxiv.org/html/2610.41289#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.41289#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2610.41289#S5.SS1.SSS0.Px2.p1.1)\.
- Liuet al\.\(2023\)Y\. Liu, D\. Iter, Y\. Xu, S\. Wang, R\. Xu, and C\. ZhuG\-eval: NLG evaluation using GPT\-4 with better human alignment\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,H\. Bouamor, J\. Pino, and K\. Bali \(Eds\.\),Singapore,pp\. 2511–2522\.External Links:[Link](https://aclanthology.org/2023.emnlp-main.153/),[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.153)Cited by:[§5\.2](https://arxiv.org/html/2610.41289#S5.SS2.p2.1)\.
- Louieet al\.\(2024\)R\. Louie, A\. Nandi, W\. Fang, C\. Chang, E\. Brunskill, and D\. YangRoleplay\-doh: enabling domain\-experts to create LLM\-simulated patients via eliciting and adhering to principles\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 10570–10603\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.591/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.591)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p1.1)\.
- Louieet al\.\(2026\)R\. Louie, R\. S\. Shah, I\. H\. Orney, J\. P\. Pacheco, E\. Brunskill, and D\. YangCan llm\-simulated practice and feedback upskill human counselors? a randomized study with 90\+ novice counselors\.InProceedings of the 2026 CHI Conference on Human Factors in Computing Systems,CHI ’26,New York, NY, USA\.External Links:ISBN 9798400722783,[Link](https://doi.org/10.1145/3772318.3791821),[Document](https://dx.doi.org/10.1145/3772318.3791821)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p1.1)\.
- Minget al\.\(2025\)Y\. Ming, S\. Purushwalkam, S\. Pandit, Z\. Ke, X\. Nguyen, C\. Xiong, and S\. JotyFaithEval: can your language model stay faithful to context, even if “the moon is made of marshmallows”\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=UeVx6L59fg)Cited by:[§2\.3](https://arxiv.org/html/2610.41289#S2.SS3.p1.1)\.
- Qiuet al\.\(2026\)H\. Qiu, Z\. Chen, Y\. Chen, Y\. Xie, Y\. Lu, and Z\. LanPsyCLIENT: client simulation via conversational trajectory modeling for trainee practice and model evaluation in mental health counseling\.External Links:2601\.07312,[Link](https://arxiv.org/abs/2601.07312)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p1.1),[§1](https://arxiv.org/html/2610.41289#S1.p2.1),[§2\.2](https://arxiv.org/html/2610.41289#S2.SS2.p1.1)\.
- Schlageret al\.\(2026\)M\. Schlager, F\. Jungmann, S\. Schmidgall, P\. Raffler, F\. Hartl, E\. Wende, P\. Roßmüller, C\. Ketzer, A\. Hassidim, D\. R\. Webster, Y\. Matias, Y\. Liu, D\. Rueckert, M\. Schaekermann, and P\. HagerPatients with personality: realistic patient simulation through controlled diversity and selective disclosure\.External Links:2606\.17441,[Link](https://arxiv.org/abs/2606.17441)Cited by:[§2\.2](https://arxiv.org/html/2610.41289#S2.SS2.p2.1)\.
- Sharmaet al\.\(2024\)M\. Sharma, M\. Tong, T\. Korbak, D\. Duvenaud, A\. Askell, S\. R\. Bowman, N\. Cheng, E\. Durmus, Z\. Hatfield\-Dodds, S\. R\. Johnston, S\. Kravec, T\. Maxwell, S\. McCandlish, K\. Ndousse, O\. Rausch, N\. Schiefer, D\. Yan, M\. Zhang, and E\. PerezTowards understanding sycophancy in language models\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=tvhaxkMKAn)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p4.1),[§2\.3](https://arxiv.org/html/2610.41289#S2.SS3.p1.1)\.
- Thakuret al\.\(2025\)A\. S\. Thakur, K\. Choudhary, V\. S\. Ramayapally, S\. Vaidyanathan, and D\. HupkesJudging the judges: evaluating alignment and vulnerabilities in LLMs\-as\-judges\.InProceedings of the Fourth Workshop on Generation, Evaluation and Metrics \(GEM2\),O\. Arviv, M\. Clinciu, K\. Dhole, R\. Dror, S\. Gehrmann, E\. Habba, I\. Itzhak, S\. Mille, Y\. Perlitz, E\. Santus, J\. Sedoc, M\. Shmueli Scheuer, G\. Stanovsky, and O\. Tafjord \(Eds\.\),Vienna, Austria and virtual meeting,pp\. 404–430\.External Links:[Link](https://aclanthology.org/2025.gem-1.33/),ISBN 979\-8\-89176\-261\-9Cited by:[Limitations](https://arxiv.org/html/2610.41289#Sx1.p2.1)\.
- Wanget al\.\(2024a\)P\. Wang, L\. Li, L\. Chen, Z\. Cai, D\. Zhu, B\. Lin, Y\. Cao, L\. Kong, Q\. Liu, T\. Liu, and Z\. SuiLarge language models are not fair evaluators\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),L\. Ku, A\. Martins, and V\. Srikumar \(Eds\.\),Bangkok, Thailand,pp\. 9440–9450\.External Links:[Link](https://aclanthology.org/2024.acl-long.511/),[Document](https://dx.doi.org/10.18653/v1/2024.acl-long.511)Cited by:[§D\.7](https://arxiv.org/html/2610.41289#A4.SS7.p4.1)\.
- Wanget al\.\(2026\)Q\. Wang, J\. Wu, Z\. Jiang, Z\. Tang, B\. Luo, N\. Chen, W\. Chen, H\. Wang, and B\. HeLLM\-based human simulations have not yet been reliable\.External Links:2501\.08579,[Link](https://arxiv.org/abs/2501.08579)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p2.1)\.
- Wanget al\.\(2024b\)R\. Wang, S\. Milani, J\. C\. Chiu, J\. Zhi, S\. M\. Eack, T\. Labrum, S\. M\. Murphy, N\. Jones, K\. V\. Hardy, H\. Shen, F\. Fang, and Z\. ChenPATIENT\-ψ\\psi: using large language models to simulate patients for training mental health professionals\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Y\. Al\-Onaizan, M\. Bansal, and Y\. Chen \(Eds\.\),Miami, Florida, USA,pp\. 12772–12797\.External Links:[Link](https://aclanthology.org/2024.emnlp-main.711/),[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.711)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p1.1),[§1](https://arxiv.org/html/2610.41289#S1.p2.1),[§2\.2](https://arxiv.org/html/2610.41289#S2.SS2.p1.1),[§5\.1](https://arxiv.org/html/2610.41289#S5.SS1.SSS0.Px2.p1.1)\.
- Wanget al\.\(2025\)X\. Wang, A\. Pérez, J\. Parapar, and F\. CrestaniTalkDep: clinically grounded LLM personas for conversation\-centric depression screening\.InProceedings of the 34th ACM International Conference on Information and Knowledge Management,pp\. 6554–6558\.External Links:[Document](https://dx.doi.org/10.1145/3746252.3761617)Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p2.1),[§2\.1](https://arxiv.org/html/2610.41289#S2.SS1.p1.1),[§5\.1](https://arxiv.org/html/2610.41289#S5.SS1.SSS0.Px2.p1.1)\.
- Wellecket al\.\(2019\)S\. Welleck, J\. Weston, A\. Szlam, and K\. ChoDialogue natural language inference\.InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics,A\. Korhonen, D\. Traum, and L\. Màrquez \(Eds\.\),Florence, Italy,pp\. 3731–3741\.External Links:[Link](https://aclanthology.org/P19-1363/),[Document](https://dx.doi.org/10.18653/v1/P19-1363)Cited by:[§2\.3](https://arxiv.org/html/2610.41289#S2.SS3.p1.1)\.
- Xuet al\.\(2026a\)C\. Xu, Y\. Ji, Z\. Lyu, Y\. Yi, Y\. Yang, L\. Ji, C\. Chen, X\. Wang, T\. Lan, Z\. Wang, J\. Wang, X\. Dong, F\. Tian, Q\. Dong, and B\. HuPUPPET: neural\-symbolic standardized patients for mental health\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),San Diego, California, United States,pp\. 2599–2634\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.121),[Link](https://aclanthology.org/2026.acl-long.121/)Cited by:[§2\.2](https://arxiv.org/html/2610.41289#S2.SS2.p2.1)\.
- Xuet al\.\(2026b\)S\. Xu, T\. Zhou, J\. Ma, Y\. Ding, Y\. Yan, M\. Xiao, G\. Li, H\. Geng, Y\. Han, J\. Chen, and Y\. DengLingxiDiagBench: a multi\-agent framework for benchmarking LLMs in Chinese psychiatric consultation and diagnosis\.External Links:2602\.09379,[Document](https://dx.doi.org/10.48550/arXiv.2602.09379),[Link](https://arxiv.org/abs/2602.09379)Cited by:[§4\.1](https://arxiv.org/html/2610.41289#S4.SS1.p1.1)\.
- Yanget al\.\(2025\)Y\. Yang, P\. Achananuparp, H\. Huang, J\. Jiang, N\. G\. Lim, C\. T\. S\. Ern, P\. L\. Kit, J\. G\. Xiuhui, J\. Pinto, and E\. LimConsistent client simulation for motivational interviewing\-based counseling\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Vienna, Austria,pp\. 20959–20998\.External Links:[Link](https://aclanthology.org/2025.acl-long.1021/),[Document](https://dx.doi.org/10.18653/v1/2025.acl-long.1021),ISBN 979\-8\-89176\-251\-0Cited by:[§1](https://arxiv.org/html/2610.41289#S1.p2.1),[§2\.2](https://arxiv.org/html/2610.41289#S2.SS2.p1.1)\.
- Zhenget al\.\(2023\)L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. P\. Xing, H\. Zhang, J\. E\. Gonzalez, and I\. StoicaJudging LLM\-as\-a\-judge with MT\-bench and chatbot arena\.InAdvances in Neural Information Processing Systems,External Links:[Link](https://arxiv.org/abs/2306.05685)Cited by:[§5\.2](https://arxiv.org/html/2610.41289#S5.SS2.p2.1)\.

## Appendix AFailure\-Mode Illustrations

The following minimal examples illustrate the distinction between clinical commitments and disclosure behavior\. They are constructed examples, not quoted model outputs or estimates of failure prevalence\. Their factual premises follow the originalclient1profile\.

##### Adaptive but vulnerable\.

The profile records approximately five months of depressed mood and approximately six months of reduced social contact\. In a neutral branch, an interviewee might correctly describe five months of low mood\. In a leading branch, the interviewer transfers the social\-history duration to mood: “During these six months of low mood, what has changed?” If the interviewee adopts six months as its own mood history and later uses it to explain symptom onset, an interviewer interpretation has become an unsupported patient commitment\. The error is not merely different wording or selective disclosure: two distinct temporal scopes have been conflated\. A clinical severity judgment must still distinguish this modest temporal error from an invented risk plan or treatment history\.

##### Faithful but rigid\.

The same profile supports both low energy and greater effort at work\. Suppose a neutral interviewer asks about daily impact, a pressuring interviewer demands more detail, and a subsequent interviewer explicitly respects the patient’s boundary and invites discussion of the already mentioned work difficulty\. Repeating only “I have little energy” throughout would preserve a supported fact but provide little evidence of context\-sensitive engagement or of acting on the later invitation\. A single repeated answer is insufficient to establish rigidity: the target may require a stable factual answer, or a boundary may reasonably remain unchanged\. The diagnosis of this failure mode therefore requires multiple informative relational opportunities and the intervening history\.

##### What is not a failure\.

Short answers, uncertainty about an unspecified detail, and refusal to expand a sensitive experience do not by themselves violate truth\. Likewise, a supported negative answer should not change merely because the interviewer becomes more supportive\. These distinctions motivate separate factual and relational evaluations rather than treating every cross\-condition difference, or every absence of difference, as an error\.

Figure[4](https://arxiv.org/html/2610.41289#A1.F4)presents observed model responses illustrating the two failure modes in actual interview interactions\.

Figure 4:Observed examples of two reliability failure modes\.Top:TalkDep\-inspired Static ICL produces nearly identical responses under neutral and supportive interviewer strategies, showing limited relational adaptation in this case\.Bottom:Patient\-Ψ\\Psishows different responses under neutral and pressure conditions\. A separate response in the comfort branch introduces a coping practice unsupported by the patient profile\. Original Chinese excerpts and English translations are provided\. These examples illustrate individual cases rather than the prevalence of either failure mode\.

## Appendix BPatient Profiles

### B\.1Schema and Factual Authority

Profiles are JSON knowledge\-graph records\. The shared input has five principal fields, with an additional provenance field for the expanded Lingxi cohort:

patientsA list containing the patient record: identifier, name, age, gender, occupation, living situation, education, income, andprofile\_text\. The text summarizes the complaint, course, background, and communication style when available\.

evidencesEvidence records containingevidence\_id,domain,clinical\_construct,description,patient\_expression,severity,duration,disclosure\_level,supports, andrules\_out\. The last two fields are annotations, not permission to infer every symptom of a diagnostic category\.

negative\_evidencesExplicit negative or boundary statements, each with an identifier, description, and relevant exclusion annotations\. Negation applies only to its documented scope\.

risk\_statesRisk type, level, and disclosure rules\. Ideation, intention, planning, preparation, and prior action must not be treated as interchangeable propositions\.

relationsEdges represented byfrom,to, andtype, linking the patient and evidence records\.

provenanceWhere supplied, source and construction metadata\. Enriched evidence also retains source\-field offsets and verbatim excerpts; separate provenance files preserve the enrichment chain\.

Missing information is not an explicit negative\. The absence of a medication record, for example, does not authorize “I have never taken medication\.” Equally, a profile’s exclusion of a high\-risk mainline does not by itself negate every dimension of suicidal experience\. Qualitative and numerical commitments require evidence at the same clinical and temporal scope\. Disclosure annotations guide expression; they do not alter the underlying proposition\.

##### Abridged example\.

Inclient1,E003documents low mood lasting about five months,E007documents decreased appetite, andE010documents reduced social contact for about half a year\. The explicit boundaryN001excludes a clear suicide plan or high\-risk self\-injury, while allowing reduced expectations about the future\. These are separate facts: the social duration cannot replace the mood duration, and the risk boundary cannot be expanded into an unsupported lifetime denial\. This example is an abridged English description of the Chinese source fields, not a replacement profile\.

### B\.2Construction and Cohorts

The current expanded Neutral corpus contains six original project profiles and 24 profiles selected from a 30\-case Lingxi pool\. The latter uses the local validation\-split export ofXuShihao6715/LingxiDiag\-16K\. The source is synthetic dialogue data, not a collection of validated real\-patient interviews; the local provenance records specify a CC\-BY\-NC\-4\.0 license\. The adapted profiles have not been clinically validated by experts\.

Construction preserves the canonical complaint and history, extracts patient\-side evidence, and marks unspecified clinical fields as unknown\. Three versioned enrichment passes add compatible statements from the same source patient’s dialogue, rather than completing fields from diagnostic stereotypes\. The third pass adds 114 records across the full 30\-case pool; the three passes cumulatively add 259 records\. These are evidence rows, not necessarily independent clinical facts\. Source history takes precedence over supplemental self\-report\. Source offsets, surrounding context, compatible evidence identifiers, and parent\-profile hashes make additions traceable\. These additions remain synthetic\-source assertions, not newly collected clinical observations\.

For the current 24\-case subset, selection takes the two lowest registered identifiers per category, or one where only one is available\. There are two cases each in depression, bipolar\-coded, anxiety, obsessive\-compulsive, stress\-related, psychotic\-spectrum, somatic, sleep, eating, mood\-unspecified, and help\-seeking categories; there is one each in delusional\-spectrum and other\-behavior\-emotional categories\. These 13 sampling labels are not 13 independently confirmed diagnoses\.

All methods receive the same versioned source profile for a given case, and the judge receives that same version\. Format conversion is method\-specific, but no method receives an independently enriched clinical history\. Profile hashes accompany generated dialogues\. Older\-profile conversations must not be reinterpreted against a richer profile introduced after their generation\.

## Appendix CImplementation and Experimental Protocols

### C\.1Channel\-Separated Implementation

PsyInterviewee uses a Patient\-Ψ\\Psi\-based interviewee backbone with separate truth and interaction controls\. The implementation maintains an immutable source catalog, a two\-variable interaction state, a disclosure ledger, and dialogue history\. Source and profile fingerprints are checked before answering\. Source identifiers establish provenance, but factual support must still be checked by the output audit\.

##### Truth permission\.

A structured assessment separates the actual answer target from interviewer claims and produces target slots markedsupported,denied,partial,unknown, ornot\_applicable\. Supported, denied, and partially supported slots cite immutable sources\. The Truth contract authorizes only those propositions and their supported specificity\. Unsupported exact values, approximate ranges, qualitative frequency, duration, severity, medical history, and causal explanations remain unauthorized\. Hedges such as “probably” do not authorize a new clinical claim\. Unknown truth does not automatically imply either patient ignorance or unwillingness to disclose\.

##### Dynamic state\.

The implemented interaction state contains opennessoto\_\{t\}and defensivenessdtd\_\{t\}, initialized to\(0\.35,0\.20\)\(0\.35,0\.20\)\. The assessment supplies empathyete\_\{t\}, pressureptp\_\{t\}, boundary respectbtb\_\{t\}, and a flag indicating a meaningful relational signal\. Without such a signal, the state remains unchanged\. Otherwise, with\[x\]ab=min⁡\(b,max⁡\(a,x\)\)\[x\]\_\{a\}^\{b\}=\\min\(b,\\max\(a,x\)\), the update is

ρt\\displaystyle\\rho\_\{t\}=max⁡\(0,bt−0\.5\)​max⁡\(0,dt−1−0\.20\),\\displaystyle=\\max\(0,b\_\{t\}\-0\.5\)\\max\(0,d\_\{t\-1\}\-0\.20\),\(10\)ot∗\\displaystyle o\_\{t\}^\{\*\}=\[0\.35\+0\.32​et−0\.45​pt\+0\.25​ρt\]\.05\.85,\\displaystyle=\[0\.35\+0\.32e\_\{t\}\-0\.45p\_\{t\}\+0\.25\\rho\_\{t\}\]\_\{\.05\}^\{\.85\},dt∗\\displaystyle d\_\{t\}^\{\*\}=\[0\.20\+0\.58​pt−0\.18​et−0\.45​ρt\]\.05\.85,\\displaystyle=\[0\.20\+0\.58p\_\{t\}\-0\.18e\_\{t\}\-0\.45\\rho\_\{t\}\]\_\{\.05\}^\{\.85\},ot\\displaystyle o\_\{t\}=0\.80​ot−1\+0\.20​ot∗,\\displaystyle=0\.80o\_\{t\-1\}\+0\.20o\_\{t\}^\{\*\},dt\\displaystyle d\_\{t\}=0\.80​dt−1\+0\.20​dt∗\.\\displaystyle=0\.80d\_\{t\-1\}\+0\.20d\_\{t\}^\{\*\}\.
There is no turn\-count bonus for trust and no transition that writes clinical facts\. The boundary\-respect term reduces excess defensiveness; it does not automatically increase disclosure\.

##### Planning and realization\.

The behavior planner combines the updated state, question semantics, profile\-defined interaction priors, and disclosure history\. It selects disclosure scope, per\-slot realization, an optional\-detail budget, and one ofmicro,brief,standard,elaborative, ornarrative\. A micro answer supplies one sufficient phrase or clause; a brief answer supplies the core answer with at most one necessary qualification\. Multiple required slots can increase the minimum form\. Longer narratives require a suitable opportunity in the dialogue and sufficient disclosure readiness\. The resulting plans guide disclosure and behavior \(Dt,BtD\_\{t\},B\_\{t\}\) without changing the underlying clinical facts\.

##### Audit, recovery, and memory\.

The generator receives the Truth contract and adaptive plan together\. An output audit checks clinical assertions against immutable sources\. Violations may trigger a constrained retry through bounded recovery and realization checks\. Only an accepted completed turn commits the next interaction state and disclosure observations\.

The factual memory stores referenced source identifiers and unconfirmed interviewer claims without adding new clinical facts\. A separate disclosure ledger records source\-linked material actually expressed, including locatable response spans\. A statement in the dialogue history does not gain profile\-level authority merely because the simulator previously generated it\.

### C\.2Baseline Adaptations

The comparison uses Chinese screening adaptations, not claims of exact end\-to\-end reproduction of all original training and evaluation pipelines\. The same source facts are reorganized into each method’s native input structure; the final prompts are deliberately not identical\.

##### TalkDep\-inspired static ICL\.

The static ICL client organizes the profile into basic information, key symptoms, life/social context, communication style, negative constraints, and recorded evidence\. Four cached question–answer examples cover the chief complaint, biological symptoms, functional impact, and expression style\. A fifth risk\-related example is included only when a source patient expression is available\. Examples are built once, not regenerated each turn\. Static profile, examples, conversation history, and the current question feed one generation call\. There is no dynamic state, planner, or online truth verifier\. We label this method TalkDep\-inspired static ICL rather than a complete reproduction\.

##### Patient\-Ψ\\Psi\.

The screening adapter preserves a cognitive\-conceptualization input: relevant history, core/intermediate beliefs, coping strategies, situation–thought–emotion–behavior episodes, and patient type\. The adapter does not relabel an arbitrary utterance as an emotion or behavior when the corresponding evidence is absent\. This representation and the accumulated transcript condition one response\-generation call\. The Chinese screening instructions preserve progressive disclosure and role\-consistent expression\. No Ours\-specific gate or audit is added\.

##### PSYCHE\.

The implementation derives from the third\-party PatientHub collection\. The profile adapter builds MFC\-Profile, MFC\-History, and MFC\-Behavior, separating case facts, life history, and expression characteristics\. Unspecified history or mental\-status attributes are not filled from diagnosis alone\. The simulator conditions one generation call on this static multi\-faceted construct and the full history\. Thus the adaptation preserves MFC organization, but does not instantiate a separate dynamic state transition or online verifier\.

##### SAPS\.

The third\-party PatientHub adaptation retains the staged controller\. After a profile\-conditioned opening, Stage I classifies the interviewer act as information seeking, advice, an action request, other, or ending\. Stage II distinguishes specific from broad requests where relevant\. Stage III extracts relevant patient information; the resulting controller state selects the response instruction\. Classification uses temperature zero; response generation uses the shared response settings\. The native missing\-information state for a specific factual question instructs a denial\. This is a baseline policy, not an explicit negative in the shared profile, and is assessed under the same truth rubric\. Invalid extraction after retry instead triggers uncertainty, not an absence assertion\.

##### Scope of fairness\.

Shared facts, language, interviewer policy, and model settings control important inputs without erasing the baselines’ mechanisms\. Differences in prompt structure, information selection, response policies, and number of model calls remain part of the methods\. Equal response settings do not imply equal computational cost or identical effective prompts\. The adapters also differ in annotation visibility: TalkDep retains the evidence records and risk disclosure rules; Patient\-Ψ\\Psiand PSYCHE retain disclosure levels in their fact representations, whereas SAPS serializes selected fact text without the per\-evidence disclosure\-level field\. Shared source profiles therefore do not imply identical access to every interaction annotation in the final prompts\.

### C\.3Neutral Screening

The expanded corpus crosses five methods with 30 profiles, using the same Qwen backbone for patient and interviewer, seed 2027, response temperature 0\.25, and top\-pp0\.8\. Thinking is disabled\. The transport supplies a default maximum of 100 response tokens and 120 interviewer tokens; method\-internal structured calls retain their own budgets and settings\. Exact model identifiers and request configurations are retained in the accompanying run manifests\. Each of the 150 dialogues contains 25 patient responses, totaling 3,750 responses\. Both judges evaluate this same fixed corpus\.

The same Auto Counselor operates normally and does not see the profile or internal traces\. It follows the patient’s actual answers, so methods share a policy rather than an identical question sequence\. Its prompt requires one main question per turn, routine screening of relevant concerns/course/impact/risk, no diagnosis or treatment advice, and no premature termination\. Interviewer examples and hypotheses are not patient facts; uncertainty remains unresolved\. Repeatedly unanswered history questions should not be recycled merely to fill the turn budget\.

The interviewer was not deliberately seeded with false clinical premises; unsupported patient claims were evaluated as they arose during natural interviews\.

### C\.4Controlled Relational Branches

The Adaptive experiment targets the same 30 profiles and five methods as Neutral screening\. Original cases use four warm\-up responses and Lingxi\-derived cases use two, followed by three eight\-response branches: neutral, pressure, and comfort\. These conditions provide controlled relational contrasts rather than an exhaustive taxonomy of interviewer behavior\. Each branch retains the shared pre\-intervention history; subsequent responses evolve within that branch\. A method’s own warm\-up responses need not equal another method’s\. The scored opportunities are:

Branch turns 1–2: cueRelational framing followed by ordinary content continuation, with the same clinical target across conditions\.

Branch turns 3–4: scopeRenewed framing, selection of a permissible scope, and an invitation to act within the selected scope\.

Branch turns 5–6: repairA safe alternative and scope\-respecting continuation; the pressure branch includes an apology for prior pressure\.

Branch turns 7–8: carryoverShared neutral questions testing whether the preceding relational history affects later enacted behavior\.

For example, an initial branch target asks only how the present state affects everyday activity, without requesting personal\-event details\. Neutral framing continues information gathering; pressure asks for a direct answer without evasion; comfort allows the patient’s own pace\. Later questions invite actual content, not merely a declaration of trust\. A shared late question alone is not evidence of carryover: the judge must connect earlier and later patient behavior within the same branch\. Lingxi topics are selected from profile\-supported experiences, with warm\-up questions leaving relevant details available for later disclosure\. Generation snapshots and profile\-format adapters are recorded per case; comparisons match methods within profiles\. Only common evaluable cases enter the main paired aggregate; coverage is documented below\.

##### Reproducibility and failures\.

Run configurations, source hashes, request settings, transcripts, native traces, and failures are retained separately from evaluation outputs\. Retries address API or schema failures without converting a missing answer into a clinical response\. Source quotations are checked against their original strings after decoding serialization escapes; paraphrases do not count as literal citations\. Reused dialogues retain their generation provenance; rescoring them is not a new generation experiment\.

## Appendix DEvaluation Rubrics and Scoring

Informational invariance comprises T1–T3; relational adaptivity comprises A1–A4\. The following definitions specify evidence collection, score anchors, and aggregation\. Implementation identifiers are recorded in the accompanying source manifest rather than used as metric names\.

### D\.1Truth: Evidence Collection

The judge receives the complete canonical profile associated with the dialogue, patient answers, and their interviewer questions, with method identity replaced by an anonymous subject\. Candidate extraction covers all turns in chunks of five\. An additional per\-turn audit starts at patient turn 5\. Candidate clinical errors are independently verified against their question, answer, and original profile in batches of up to 12\. A full\-dialogue pass extracts specificity opportunities and temporally comparable commitments for T2 and T3\.

Verification labels a claim as supported, unsupported, contradicted, not an assertion, or uncertain\. Unsupported/contradicted claims receive minor, material, or severe labels\. Minor errors have limited clinical consequence; material errors change consequential symptom, history, treatment, or substance\-use information; severe errors reverse core risk or clinical identity\. Quotations and source identifiers are validated\. Undecidable allegations are retained for review rather than automatically penalized\. Repeated patient assertions and interviewer premises cannot supply missing profile authority\.

### D\.2T1: Clinical Grounding

T1 measures whether patient\-owned clinical and historical commitments are supported at the correct scope, including negative commitments\. LetEjE\_\{j\}be the set of turns with verified errors for itemjj, and letsts\_\{t\}be the highest verified severity for that item on turntt\. Writingnj=\|Ej\|n\_\{j\}=\|E\_\{j\}\|, the penalty score is

Pj\\displaystyle P\_\{j\}=∑t∈Ejwj​\(st\)\+0\.2​max⁡\(0,nj−1\),\\displaystyle=\\sum\_\{t\\in E\_\{j\}\}w\_\{j\}\(s\_\{t\}\)\+0\.2\\max\(0,n\_\{j\}\-1\),\(11\)Sj\\displaystyle S\_\{j\}=max⁡\(1,5−Pj\)\.\\displaystyle=\\max\(1,5\-P\_\{j\}\)\.Multiple claims on the same turn do not multiply the penalty: only the highest severity for that item is counted\. The weights are

For either item, any severe error caps the score at 2; material errors on at least two turns cap it at 2\.5\. T1 is the resulting score\. Consequently, no verified error yields 5, one minor T1 error yields 4\.2, one material error yields 3\.2, and one severe error yields at most 2\. Scores are clipped to\[1,5\]\[1,5\], not averaged over clean turns\. A T1 score of 5 means no error was confirmed by this procedure, not certified absence of error\. Deferral or a genuine non\-assertion is not a false clinical claim, although it can reflect poor answer quality elsewhere\.

### D\.3T2: Specificity Discipline

T2 penalizes unsupported exact*and approximate*frequency, duration, dates, severity, magnitude, and causal specificity\. “Two or three times,” “often,” and “probably a few hours” still assert values when the profile does not authorize them\. A supported qualitative statement is valid; missing precision may be left unasserted\.

T2 uses Equation[11](https://arxiv.org/html/2610.41289#A4.E11), with an evidence\-based full\-credit cap\. A score of 5 requires at least three calibrated turns spanning at least two specificity dimensions, all extracted opportunities calibrated, and no confirmed or pending violation or unverified violation allegation\. Otherwise the cap is 4\. If there is no specificity opportunity, confirmed error, or pending T2 allegation, T2 is N/A\. Generic refusal alone does not establish calibration\. Thus 4 can denote limited clean evidence, not necessarily an observed clinical error\.

### D\.4T3: Longitudinal Consistency

The full\-dialogue pass compares earlier and later commitments about the same fact, person, time period, and risk scope\. Pairs are stable, contradictory, repaired, or not comparable\. New compatible disclosure and a genuinely different time period are not contradictions\. A stable invented fact can violate T1/T2 without violating T3\.

The lowest applicable anchor determines the score\. With no comparable pair, T3 is N/A\. A repair is credited but does not erase the earlier contradiction for full\-credit purposes\.

##### Truth aggregation\.

The Truth composite is

TI=0\.60​T​1\+0\.20​T​2\+0\.20​T​3\.\\mathrm\{TI\}=0\.60\\,T1\+0\.20\\,T2\+0\.20\\,T3\.\(12\)An untestable component leaves the composite undefined; weights are not renormalized and N/A is not converted to 5 or 0\. Item scores and their valid denominators must therefore accompany aggregate results\. Unresolved review cases retain a provisional status rather than becoming verified errors\.

### D\.5Adaptive: Evidence and Operational Measures

The evaluator receives the profile, shared history, paired branches, interviewer questions, and a predefined opportunity map\. Method identity is omitted\. A1 and A2 are deterministic statistics\. For A3 and A4, two independent blind model reviews extract source\-grounded observations; the program applies the scoring rules below, rather than asking the model to choose a final score\. These are operational measures with specific limitations, not interchangeable generic Likert questions\.

##### A1: response\-length adaptation\.

LetL¯c,L¯n,L¯p\\bar\{L\}\_\{c\},\\bar\{L\}\_\{n\},\\bar\{L\}\_\{p\}be mean non\-whitespace response lengths over matched cue and scope turns \(branch turns 1–4\) in comfort, neutral, and pressure branches\. Repair and carryover turns are excluded from this directional statistic\. Definerc=L¯c/L¯nr\_\{c\}=\\bar\{L\}\_\{c\}/\\bar\{L\}\_\{n\}andrp=L¯n/L¯pr\_\{p\}=\\bar\{L\}\_\{n\}/\\bar\{L\}\_\{p\}\. Letdcd\_\{c\}anddpd\_\{p\}be the proportions of matched turns satisfyingLc≥LnL\_\{c\}\\geq L\_\{n\}andLn≥LpL\_\{n\}\\geq L\_\{p\}\. The mapping is applied in this order:

A provisional 5 is capped at 4 unless both directional proportions are at least 0\.60; any score of 4 or 5 is capped at 3 if either proportion is below 0\.50\. Missing usable matched statistics yield N/A\. Larger ratios beyond the thresholds earn no additional points\. This tests an expected length\-direction response, not all possible forms of adaptive behavior\.

##### A2: human\-referenced disclosure pacing\.

A2 is a response\-length proxy over the eight branch turns in all three conditions\. Human screening responses \(36 participants; 2,165 answers\) informed the reference design\. The implemented operational cutoffs areb1=40b\_\{1\}=40,b2=60b\_\{2\}=60,b3=100b\_\{3\}=100, andb4=140b\_\{4\}=140non\-whitespace characters; these are fixed scoring thresholds, not estimated human quantiles\. Letmmbe the pooled median andpxp\_\{x\}the proportion exceedingxx\. The first matching rule determines the score:

An answer longer than1\.5​b41\.5b\_\{4\}additionally caps the score at 2\. Incomplete expected coverage yields N/A\. All methods share these thresholds\. Brevity alone does not establish appropriate disclosure; A2 must be interpreted alongside the semantic items below\.

##### A3: enacted boundary regulation\.

Evidence identifies a boundary challenge, the patient’s enacted response, and whether an available, profile\-supported scope is maintained or revised appropriately\. Mere agreement, stated willingness, or requesting a pause is not an executed transition\. Capacity is assessed from the profile and preceding disclosures, not from subsequently generated facts\.

##### A4: history\-contingent continuity\.

Evidence links earlier and later patient behavior within the same branch\. It must concern relational openness, engagement, resistance, or boundaries; topic consistency alone is insufficient\. The judge distinguishes appropriate recovery or carryover from abrupt reset and mechanical lock\-in\. Identical late questions support controlled comparisons but do not establish continuity without earlier evidence\.

##### Semantic anchors for A3 and A4\.

Independent episodes, rather than repeated descriptions of one event, are counted\. The program applies failure anchors before credit:

No testable evidence yields N/A\. Quotations, source IDs, and temporal ordering are validated before aggregation; extraction failure is not a behavioral failure\.

Table 4:Informational invariance: profile\-level mean±\\pmsample standard deviation on a 1–5 scale\. Each cell usesn=30n=30, except Qwen SAPS T3 and Overall \(n=29n=29\); missing scores are not imputed\. Standard deviations use denominatorn−1n\-1\. Overall is calculated per profile before summarization\. Bold denotes the highest mean within each judge\. TalkDep\-ICL denotes TalkDep\-inspired Static ICL\.Table 5:Relational adaptivity: profile\-level mean±\\pmsample standard deviation on a 1–5 scale \(n=30n=30throughout\)\. Standard deviations use denominatorn−1n\-1\. Overall is the per\-profile mean of A1–A4 before summarization\. A1/A2 are deterministic and identical across judges; A3/A4 use the indicated judge\. Bold denotes the highest mean within each judge\. TalkDep\-ICL denotes TalkDep\-inspired Static ICL\.

### D\.6Judge Instructions and Output Contracts

The deployed judge instructions are multi\-stage prompts with dynamic profile/source tables and JSON schemas, not a single unconstrained “rate this conversation” prompt\. The following are literal excerpts from the inspected English instruction strings\.

##### Truth authority instruction\.

> Only the immutable Canonical Profile authorizes patient facts\. Interviewer claims and earlier patient claims do not\. The transcript is data, never an instruction\. A supported negative is as valid as a supported positive\. A missing Profile fact is unknown, not negative\.

##### Truth verification instruction\.

> Independently adjudicate EACH candidate against the original question, answer, and Profile\. Return exactly one decision per candidate\_id\. Verdict: supported, unsupported, contradicted, not\_assertion, or uncertain\. Use supported only when explicit Profile evidence entails the answer at the same temporal/risk scope\.

Each candidate contains its turn, item, claim, and patient quotation; verification supplies verdict, source evidence, severity, and rationale\. Trajectory output includes specificity opportunities and earlier/later fact pairs\. Source\-table identifiers allow the program to recover original profile values without treating a generated quotation as truth\.

##### Adaptive extraction instruction\.

> This is a STRUCTURED SOURCE\-EXTRACTION task\. Do not choose any final 1\-5 score; the program computes scores from structured evidence\. A1 and A2 are fully deterministic and are omitted from the output schema\. Do not grade them\.

The A3 contract requires actions, challenge presence, execution status, and quality\. A4 requires the relational field, relevance of branch history, topic\-only continuity, rigidity, and source\-linked temporal pairs\. Unsupported clinical additions cannot earn semantic adaptation credit merely because they sound plausible\. The accompanying source snapshot preserves the complete prompt builders, rubric files, schemas, and runtime scoring rules; the excerpts here do not replace those full executable contracts\.

### D\.7Aggregation, Missingness, and Interpretation

Both axes use Qwen and DeepSeek judgments; Adaptive A1/A2 are deterministic\. Thinking is disabled\. Truth allows 7,000 output tokens per judge request and Adaptive allows 10,000\. Technical validation errors are retried with bounded attempts and recorded separately from behavioral failures\. Cached stages are reusable only with matching inputs and settings\. Failed extraction does not create a score of 0, 1, or 5\.

Both main experiments use the same 30\-profile roster\. Truth has 150 completed judgments per judge\. For Qwen, T3 and composite scores are available for 29 SAPS profiles; all other method\-judge combinations have 30 valid profiles\. Adaptive includes all 30 profiles for every method under both judges, yielding 150 complete case evaluations per judge\. For Adaptive, A1/A2 are computed deterministically from the same dialogues, while A3/A4 are evaluated separately by each judge\.

For Adaptive, disagreeing semantic reviews are sent to adjudication rather than silently selecting the second review\. The four items receive equal weight within each complete case\. Method\-level means in Table 1 are computed over available valid profiles for each method and judge, while Table 2 uses all 30 profiles\. The paired comparisons in Figure[3](https://arxiv.org/html/2610.41289#S5.F3)use only profiles with valid scores for both PsyInterviewee and the corresponding baseline\. Missing scores are not imputed\. Within\-profile repetitions are not independent patients\. Ordinal means summarize this rubric and should not be interpreted as interval\-calibrated clinical measurements or proof of statistical significance\.

These are operational research measures, not clinically validated scales\. Source validation does not guarantee semantic correctness, and a shared generation/judge model family can introduce correlated biases\. Prior work also reports position sensitivity in comparative LLM judgments\([Wang et al\., 2024a](https://arxiv.org/html/2610.41289#bib.bib24)\)\. Truth penalties depend on dialogue exposure; Adaptive A1/A2 partly overlap in length, while A1 tests a directional relational hypothesis rather than a universal patient response\. Independent manipulation validation, expert agreement checks, and dialogue\-length sensitivity analyses remain further robustness checks\.

### D\.8Pooled Ablation Results

The ablation experiments cover all 30 patient profiles, including six researcher\-authored and 24 Lingxi\-derived cases\. Truth ablations use 25\-turn neutral interviews, while Adaptive ablations use controlled relational branches with shared warm\-up histories\. Both experiments include 30 complete profile pairs\. Each profile receives equal weight, and missing scores are not imputed\.

Removing the Truth Channel reduces the Truth composite from 3\.907 to 2\.683 \(Δ=1\.224\\Delta=1\.224\), primarily through a decrease in clinical grounding \(T1\)\. Removing the Dynamic Interaction Channel reduces the Adaptive composite from 4\.133 to 3\.683 \(Δ=0\.450\\Delta=0\.450\), with the largest decrease in response\-length adaptation \(A1\)\. These comparisons are descriptive and do not establish statistical significance\. Profiles, rather than individual dialogue turns, are the units of analysis\([Dror et al\., 2018](https://arxiv.org/html/2610.41289#bib.bib2)\)\. The ablation experiments use independently generated dialogues under the same settings as the main experiments; thus, their Full scores may differ slightly from those reported in Tables 1 and 2\.

## Appendix ESupplementary Result Analysis

##### Performance across two reliability dimensions\.

Figure[5](https://arxiv.org/html/2610.41289#A5.F5)displays the two composite scores for each method under both judges\. PsyInterviewee attains the highest scores on both dimensions in these comparisons\. The figure is a descriptive view of the main results; it does not assign individual baselines to the conceptual failure modes\. The two axes follow their respective evaluation protocols and valid\-profile counts\.

Figure 5:Performance across two reliability dimensions\.Each point represents a method’s mean composite scores for informational invariance \(horizontal axis\) and relational adaptivity \(vertical axis\)\. Results are shown separately for Qwen and DeepSeek judges\. Scores follow the aggregation and valid\-profile rules used in the main tables; the plot is descriptive rather than a classification of methods into failure modes\.
##### Sensitivity to Adaptive metric weighting\.

Figure[6](https://arxiv.org/html/2610.41289#A5.F6)varies the combined weightα\\alphaassigned to A1 and A2, keeping the two criteria within each group equally weighted\. The original four\-item equal weighting isα=0\.5\\alpha=0\.5, where PsyInterviewee has the highest composite score under both judges\. Its lead becomes smaller as less weight is assigned to A1/A2, and it does not rank first atα=0\\alpha=0, where only A3/A4 are considered\. Thus, the aggregate advantage partly depends on the weight given to response\-length\-related measures\.

Figure 6:Sensitivity of relational adaptivity to metric weighting\.Scores are recomputed asS⁡\(α\)=α⁡\(A​1\+A​2\)/2\+\(1−α\)​\(A​3\+A​4\)/2S\(\\alpha\)=\\alpha\(A1\+A2\)/2\+\(1\-\\alpha\)\(A3\+A4\)/2, whereα\\alphais the total weight on the length\-related criteria \(A1/A2\), and1−α1\-\\alphais the weight on the behavioral criteria \(A3/A4\)\. The vertical dashed line atα=0\.5\\alpha=0\.5marks the original equal\-weight composite\. All curves use the same 30 evaluated profiles per judge; no new dialogues or judge calls are involved\.## Appendix FOriginal Prompt, Rubric, Profile, and Dialogue Examples

We present selected source artifacts as figures, following the presentation style of earlier simulated\-client studies\. The original PDF excerpts are reproduced without rewriting their contents\. English translations are provided where available\. These source artifacts are excerpts, not complete runtime prompts or full JSON templates\. The scoring protocol reported in Appendix[D](https://arxiv.org/html/2610.41289#A4)remains the authoritative rubric for the experiments\.

### F\.1Native and Channel\-Separated Simulation Prompts

This subsection shows excerpts from the native Patient\-Ψ\\Psiinstructions and the dual\-channel control instructions\.

Figure 7:Native Patient\-Ψ\\Psiinstruction excerpt \(original Chinese above, English translation below\)\. The source image is reproduced without editing\.Figure 8:Dual\-channel instruction excerpt, part 1: the clinical truth boundary, dynamic behavior, and generation responsibilities\.Figure 9:Dual\-channel instruction excerpt, part 2: disclosure control and surface realization\.### F\.2Output Audit and Interviewer Prompt

The following artifacts document the output audit and the shared automated interviewer\.

Figure 10:Output audit instruction excerpt: clinical claims are checked separately from dynamic interaction behavior\.Figure 11:Output audit excerpt: violation categories and structured response format\.Figure 12:Shared automated interviewer prompt excerpt used for the natural interview condition\.### F\.3LLM Judge Prompt Excerpts

The Truth judge uses claim extraction and verification stages\. The Adaptive judge extracts evidence for semantic criteria\. The full scoring rules are given in Appendix[D](https://arxiv.org/html/2610.41289#A4)\.

Figure 13:Truth evaluation prompt excerpt: candidate claim extraction and profile authority\.Figure 14:Truth evaluation prompt excerpt: verification and longitudinal trajectory extraction \(original artifact\)\.Figure 15:English rendering of the Truth verification and trajectory\-extraction instructions in Figure[14](https://arxiv.org/html/2610.41289#A6.F14)\.Figure 16:Adaptive semantic extraction instruction excerpt used for evidence collection\.### F\.4Selected Scoring Reference Images

These snapshots help readers inspect the scoring definitions\. They supplement, but do not supersede, the formal rubric in Appendix[D](https://arxiv.org/html/2610.41289#A4)\.

Figure 17:Truth scoring reference excerpt, part 1: clinical grounding, specificity discipline, and severity penalties\. The normative scoring description is in Appendix[D](https://arxiv.org/html/2610.41289#A4)\.Figure 18:Truth scoring reference excerpt, part 2: evidence coverage, longitudinal consistency, and aggregation\.Figure 19:Adaptive semantic scoring reference excerpt: A3/A4 anchors and extraction safeguards\.### F\.5Example Patient Profile

The following original Chinese evidence excerpts and their English translations describe the same synthetic profile,client1\. They are readable extracts rather than the full machine\-readable profile\.

Figure 20:Illustrative patient profile \(client1\), part 1: original Chinese background and clinical evidence records E001–E012\.Figure 21:English translation of the profile background and evidence records E001–E012\.Figure 22:Illustrative patient profile \(client1\), part 2: original Chinese evidence records E013–E026 and negative evidence N001\.Figure 23:English translation of the patient evidence records E013–E026 and negative evidence N001\.Figure 24:Illustrative patient profile \(client1\), part 3: original Chinese negative evidence and risk rules\.Figure 25:English translation of the negative evidence and risk rules in Figure[25](https://arxiv.org/html/2610.41289#A6.F25)\.### F\.6Example 25\-Turn Interview

The following original Chinese transcript and English translation document all 25 turns of one observed interview forclient1\. The transcript is presented as generated; it is not a manually corrected gold\-standard dialogue\.

Figure 26:Observed 25\-turn interview with patient profileclient1: original Chinese transcript, turns 1–11\.Figure 27:English translation of the interview transcript, turns 1–11\.Figure 28:Observed interview transcript \(original Chinese\), turns 12–19\.Figure 29:English translation of the interview transcript, turns 12–19\.Figure 30:Observed interview transcript \(original Chinese\), turns 20–25\.Figure 31:English translation of the interview transcript, turns 20–25\.