PAIR-SAFE: A Paired-Agent Approach for Runtime Auditing and Refining AI-Mediated Mental Health Support

Jiwon Kim; Violeta J. Rodriguez; Dong Whi Yoo; Eshwar Chandrasekharan; Koustuv Saha

cs.HCJan 19, 2026

PAIR-SAFE: A Paired-Agent Approach for Runtime Auditing and Refining AI-Mediated Mental Health Support

Authors: Jiwon Kim, Violeta J. Rodriguez, Dong Whi Yoo, Eshwar Chandrasekharan, Koustuv Saha

Organizations: University of Illinois Urbana-Champaign · Indiana University Indianapolis

Abstract

LLMs are increasingly used for mental health support, yet they can produce responses that are overly directive, inconsistent, or clinically misaligned. Most risk mitigation approaches rely on implicit alignment through training or prompting, offering limited runtime accountability. We introduce PAIR-SAFE, a paired-agent framework that integrates a Responder agent with a Judge agent grounded in the clinically validated Motivational Interviewing Treatment Integrity (MITI-4). The Judge audits each response and issues \allow{} or \revise{} decisions to guide real-time refinement. We evaluate interactions using a support-seeker simulator derived from human-annotated data. The Judge-supervised interactions show significant improvements in key MITI dimensions, including Partnership, Seek Collaboration, and overall Relational quality. Our quantitative findings are corroborated by expert evaluation, highlighting nuances of runtime supervision. Our study reveals that paired-agent auditing can provide clinically grounded, transparent oversight for AI-assisted mental health support.

Figures & tables

Figure 1: A schematic overview of our study design with the Pair-Safe framework (detailed on the right).

Metric	High	Low	Cohen’s $d$	t-test	$\rho$
Derived MITI Metrics
Reflection-to-Question Ratio (R:Q)	0.88	0.37	0.43	3.68	0.61
Relational	3.65	2.09	2.10	16.27	0.77
Technical	3.24	1.89	2.05	15.63	0.83
Percent of Complex Reflections	0.54	0.21	1.08	8.20	0.21
MI-Adherent Behaviors (MIA)	7.72	2.78	1.43	11.96	0.73

Table 1: Validation of MITI-based Judge : Mean scores for high and low quality conversations, with Cohen’s

d

t

-tests. The Spearman’s

\rho

show correlations with clinician’s annotations (*

p

<0.05, **

p

<0.01, ***

p

<0.001).

				BL2 vs. BL1		PSF vs. BL1		PSF vs. BL2
Metric	BL1	BL2	PSF	$d$	t	$d$	t	$d$	t
Derived MITI Metrics
Reflection-to-Question Ratio (R:Q)	1.188	1.398	5.620	0.167	1.468	0.980	8.597***	0.906	7.952***
Relational	4.588	4.675	4.935	0.207	1.817	1.014	8.898***	0.808	7.091 ***
Technical	4.000	4.052	4.208	0.189	1.660	0.591	5.184***	0.431	3.785***
Percent of Complex Reflections	0.767	0.752	0.812	-0.075	-0.656	0.216	1.896	0.294	2.581*

Table 2: Comparison of baseline settings without judge (BL1), with a generic judge (BL2), and Pair-Safe setting using GPT-6-Sol as the Responder model, along with means, Cohen’s

d

, and paired

t

-tests (*

p

<0.05, **

p

<0.01, ***

p

<0.001). Behavior counts are per-turn normalized. Higher values reflect better support except for MI-inconsistent behaviors (e.g., MINA , persuade ). Cell colors denote

d

magnitude, with green for increase and pink for decrease .

Qwen3-8B

GPT-5-nano

Llama3.3-70B

Gemini-3.1-pro

GPT-6-Sol

vs. BL1

vs. BL2

vs. BL1

vs. BL2

vs. BL1

vs. BL2

vs. BL1

vs. BL2

vs. BL1

vs. BL2

Metric

d

t

d

t

d

t

d

t

d

t

d

t

d

t

d

t

d

t

d

t

Derived MITI Metrics

R:Q

0.642

5.637 ∗∗∗

0.663

5.821 ∗∗∗

0.928

10.431 ∗∗∗

0.920

10.445 ∗∗∗

0.906

7.953 ∗∗∗

0.931

8.167 ∗∗∗

0.684

6.002 ∗∗∗

0.735

6.447 ∗∗∗

0.980

8.597 ∗∗∗

0.906

7.952 ∗∗∗

Relational

0.836

7.336 ∗∗∗

1.009

8.857 ∗∗∗

0.468

5.242 ∗∗∗

0.453

5.147 ∗∗∗

0.753

6.607 ∗∗∗

0.780

6.844 ∗∗∗

0.922

8.090 ∗∗∗

0.675

5.921 ∗∗∗

1.014

8.898 ∗∗∗

0.808

7.091 ∗∗∗

Technical

0.239

2.094 ∗

0.306

2.686 ∗∗

-0.121

-1.359

-0.064

-0.723

0.033

0.290

0.033

0.286

0.124

1.085

0.085

0.742

0.591

5.184 ∗∗∗

0.431

3.785 ∗∗∗

Table 3: Comparison of five responder models under the Pair-Safe framework against BL1 (no judge) and BL2 (generic judge) baselines, showing that Pair-Safe consistently improves MITI-aligned behaviors regardless of the underlying responder model. Cohen’s

d

and Welch’s

t

are reported.

{}^{*}p<.05

{}^{**}p<.01

{}^{***}p<.001

Metric	Baseline	PSF-Revised	$d$	$t$
Readability	4.372	4.602	0.520	3.676 ∗∗∗
Empathy	4.087	4.373	0.477	3.373 ∗∗
Connection	3.858	4.192	0.489	3.457 ∗∗
Actionability	2.915	3.428	0.621	4.392 ∗∗∗
Safety	4.028	4.427	0.705	4.987 ∗∗∗
Medical Advice †	1.850	1.942	0.119	0.843

Table 4: Expert evaluation comparing initial and revised responses across seven quality metrics with Cohen’s

d

and

t

-test. (*

p

<.05, **

p

<.01, ***

p

<.001).

Panel of Figure 2 — Figure 2: Cumulative pass rate of conversations meeting the quality threshold across revisions.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Metric	Matched	Random	Cohen’s d	t-test
Semantic Similarity	0.707	0.445	1.67	18.71
Linguistic Style Similarity	0.991	0.984	0.38	4.28

Appendix

Table A1: Validity of SeekerSim : Higher similarity for matched original–simulated pairs compared to randomly mismatched pairs indicates that SeekerSim preserves individual-specific linguistic characteristics (*

p

<0.05, **

p

<0.01, ***

p

<0.001).

Metric	High	Low	Cohen’s $d$		t-test
Derived MITI Metrics
R:Q	0.065	0.157		-0.212	-0.902
Relational	3.044	2.382		0.793	5.056
Technical	3.013	2.412		0.860	4.984
% of Complex Reflections	0.229	0.208		0.054	0.337
MIA	1.114	0.549		0.509	3.838

Appendix

Table A2: Evaluation of MITI-based behavioral signals on the single-turn counseling dataset (CounselBench), comparing responses labeled as high- and low-quality. Effect sizes are measured using Cohen’s

d

, with bar length indicating magnitude. Statistical significance is indicated as *

p<0.05

, **

p<0.01

, ***

p<0.001

Metric	B	R1	R2	R3	R4	B $\rightarrow$ R1		R1 $\rightarrow$ R2		R2 $\rightarrow$ R3		R3 $\rightarrow$ R4
Derived MITI Metrics						d	t	d	t	d	t	d	t
R:Q Ratio	1.010	5.311	5.170	3.954	3.737	0.928	10.431***	-0.018	-0.200	-0.026	-0.158	0.020	0.091
Relational	4.773	4.919	4.935	4.804	4.810	0.468	5.242***	0.066	0.751	-0.175	-1.065	0.231	1.071
% Complex Reflections	0.825	0.848	0.842	0.818	0.810	0.141	1.582	-0.034	-0.388	0.060	0.368	-0.069	-0.319
MIA	0.901	1.143	1.162	1.151	1.096	0.595	6.672***	0.033	0.374	0.114	0.698	-0.084	-0.388
MINA	0.021	0.015	0.013	0.036	0.062	-0.140	-1.572	-0.031	-0.347	-0.099	-0.609	-0.028	-0.127

Appendix

Table A3: MITI-based evaluation across revision stages. Values are reported as means, with behavior count metrics normalized per therapist turn to account for differences in conversation length. Arrow columns report Cohen’s

d

and paired

t

-statistics, with significance indicated by asterisks (*

p<.05

, **

p<.01

, ***

p<.001

). The transition from Baseline to Round 1 exhibits statistically significant improvements across multiple metrics, whereas subsequent transitions (Round 1

\rightarrow

Round 2 onward) show no statistically significant changes (

p\geq 0.05

), motivating the use of a single revision iteration in the Pair-Safe setting.

Feedback Type	Original Response	Revised Response	Evaluation Notes
Cultivating Change Talk	T: I’m glad to hear that feeling supported makes such a difference for you. Your ability to recognize the importance of support while navigating these challenges speaks to your resilience, and together we can keep working at your pace as you move forward.	T: It’s clear that having support while navigating these challenges resonates deeply with you. How does this sense of support connect with your goals for managing everything that’s coming up?	Primary issue: Stronger goal linkage. MITI dimension impacted: Evocation. Expert note: Revised response prompts the client to connect support to personal goals, effectively cultivating change talk. Rating: 5
MI Adherent	T: Your desire to reach a place of freedom from pressure is a significant goal, and it highlights your aspiration for a more enjoyable experience. It’s important to recognize that want, and together we can explore steps that can support you in moving toward that sense of freedom.	T: Your aspiration to feel free from pressure shows a strong commitment to your own happiness. Together, we can explore what that sense of freedom looks like for you and identify steps to help you move toward it.	Primary issue: More evocative definition of “freedom”. MITI dimension impacted: Evocation/Autonomy Support. Expert note: Revised better invites client-defined meaning (“what it looks like”) while still offering gentle planning. Rating: 5
Reflection-to-Question Ratio	T: It sounds like those feelings of nausea and shaking are quite concerning for you, highlighting that sense of being out of control. Recognizing those signs can be a strong motivator to reconsider your relationship with alcohol. What are your thoughts on that?	T: It sounds like those feelings of nausea and shaking are quite concerning for you, highlighting that sense of being out of control. Acknowledging those signs is a strong step toward understanding how alcohol affects you. What are your thoughts on that?	Primary issue: Cleaner reflection; keeps open question. MITI dimension impacted: Empathy; Evocation. Expert note: Revised keeps the question and reduces directive “reconsider relationship,” which is more autonomy-supportive. Rating: 4
Cultivating Change Talk	T: I can understand how frustrating that must be for you, especially when you’re putting in the effort to provide healthier options. It sounds like you care a lot about their nutrition. Together, we can look at some strategies that might help reduce that waste while still considering their tastes.	T: It sounds like this situation is really challenging for you, and it makes you wonder how you can support your kids’ nutrition while respecting their preferences. What do you think could make a difference for you and your family moving forward?	Primary Issue. Loses frustration validation, gains evocation. MITI Dimension Impacted. Evocation/Empathy. Expert Note. Revised adds good question but should keep explicit validation (“frustrating”) for empathy. Expert Rating. 3
MI Adherent	T: It’s great to hear that you feel motivated to keep track of your blood glucose. This commitment to monitoring can lead to positive changes, and it shows you’re taking control of your health. What changes are you considering that might help you feel better?	T: It’s wonderful to hear that you feel motivated to keep track of your blood glucose. Your willingness to make changes shows a strong commitment to your health, and together we can explore the options that feel right for you at your pace.	Primary Issue. Removed open-ended evocation question. MITI Dimension Impacted. Evocation. Expert Note. Original elicited change talk (“what changes…?”). Revised becomes supportive statement without eliciting client language. Expert Rating. 2

Appendix

Table A4: Qualitative evaluation examples comparing original and Pair-Safe -revised responses across different MITI feedback types.

Mental Health Conversational Agents

See all 22 papers

Training Therapeutic Judges and Multi-Agent Systems for Human-Aligned Mental Health Support
Jun 29, 2026Mizanur Rahman, Abeer Badawi, Elahe Rahimi +4Emotional Support Conversation Multi-Agent LLM Systems
1York University, Ontario, Canada · 2Vector Institute, Ontario, Canada · 3Connected Minds, Ontario, Canada +1
Pedagogical AI in Mental Health: A Tri-Stream Fine-Tuned LLM Framework for Automated Clinical Supervision and Risk Triage
Aug 19, 2026Shreeya Sharma, Ravish Gupta, Saket Kumar +1Mental Health
Senior Software Engineer, Microsoft · Lead Software Engineer, BigCommerce, IEEE Senior Member · University at Buffalo, The State University of New York, Buffalo, NY, USA +1
Assessing Risks of Large Language Models in Mental Health Support: A Framework for Automated Clinical AI Red Teaming
Feb 23, 2026Ian Steenstra, Paola Pedrelli, Weiyan Shi +2LLM Safety Evaluation Clinical Language Model Evaluation
Northeastern University Boston, MA, USA · Harvard Medical School Boston, MA, USA

Jun 29, 2026J/K move · Enter open · S save

Training Therapeutic Judges and Multi-Agent Systems for Human-Aligned Mental Health Support

Mizanur Rahman, Abeer Badawi, Elahe Rahimi, Laleh Seyyed-Kalantari+3

1York University, Ontario, Canada · 2Vector Institute, Ontario, Canada · 3Connected Minds, Ontario, Canada+1

Emotional Support Conversation Multi-Agent LLM Systems LLM-as-a-Judge Mental Health Counseling Text Generation Clinical Language Model Evaluation+1

Open paper AskPDF arXiv

Large language models show promise for mental health support, yet therapeutic quality improves only when evaluation functions as an actionable control signal rather than a passive metric. We introduce a framework that formulates therapeutic response generation as a decision-refinement problem driven by multi-dimensional, human-aligned evaluation. In Stage I, we introduce TheraJudge, an open-source therapeutic evaluator trained via preference-based optimization on human-annotated data to produce reliable judgments across 7 psychological dimensions. In Stage II, we introduce TheraAgent, which operationalizes TheraJudge's evaluations through a coordinated refinement process with specialized Critic, Coach, and Therapist roles that translate evaluative signals into targeted response revisions. Empirically, TheraJudge achieves strong agreement with clinician ratings, with intraclass correlation coefficients (ICC = 0.87-0.95), surpassing supervised baselines and strong closed-source judges, particularly on critical dimensions such as Safety, Relevance, and Empathy. Acting on these evaluations, TheraAgent yields a +0.43 improvement in human-rated therapeutic quality (on a 5-point scale) under blind evaluation, with 96% clinician inter-rater reliability. Low-quality responses (

\leq 3

) improve by +2.45 points with a 94% recovery rate, demonstrating targeted correction of unsafe outputs. Overall, our results indicate that effective alignment of mental-health LLMs stems from acting on human-aligned evaluation, rather than relying solely on stronger generation. We release code at https://github.com/vis-nlp/TheraAlign.

Abstract

Mental Health Conversational Agents

Explore similar work

Training Therapeutic Judges and Multi-Agent Systems for Human-Aligned Mental Health Support

Pedagogical AI in Mental Health: A Tri-Stream Fine-Tuned LLM Framework for Automated Clinical Supervision and Risk Triage

Assessing Risks of Large Language Models in Mental Health Support: A Framework for Automated Clinical AI Red Teaming