Graph2Counsel: Clinically Grounded Synthetic Counseling Dialogue Generation from Client Psychological Graphs¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/UKPLab/emnlp2026-graph2counsel
Area: Medical NLP
Keywords: Psychological Counseling Dialogue, Client Psychological Graphs, Process-Based Therapy, LLM Fine-Tuning, Multi-Turn Clinical Evaluation
TL;DR¶
Graph2Counsel generates clinically grounded synthetic counseling dialogues by conditioning LLMs on Client Psychological Graphs (CPGs) and real counselor intervention strategies, producing high-fidelity multi-turn therapeutic sessions that outperform prior synthetic datasets across expert clinical evaluations and downstream benchmarks.
Background & Motivation¶
Mental health disorders affect nearly one in seven individuals worldwide, yet access to professional counseling remains severely constrained by clinician shortages, exorbitant costs, and pervasive social stigma. While Large Language Models (LLMs) present a promising, highly accessible, and non-judgmental medium for mental health support, deploying unadapted off-the-shelf LLMs introduces grave safety risks—ranging from hallucinated medical advice and inappropriate therapeutic interventions to potential psychological harm. Mitigating these risks requires high-quality counseling dialogue datasets for alignment and fine-tuning. However, collecting large-scale real-world counseling corpora is heavily impeded by strict confidentiality regulations, clinical ethics, and the fragility of de-identification techniques, which struggle with scalability and re-identification vulnerabilities.
To circumvent this bottleneck, recent research has explored synthetic counseling session generation, producing datasets such as Psych8k, MentalChat16k, MDD-5k, and CACTUS. Nonetheless, current synthesis paradigms predominantly depend on static, unstructured, or semi-structured inputs, such as demographic personas, discrete questionnaire scores, or isolated symptom checklists. These representations treat a client's mental state as a collection of disjoint snapshots, completely overlooking the dynamic, functional dependencies among cognitive, emotional, and behavioral processes (for instance, how a fear of negative evaluation triggers repetitive overthinking, or how mindfulness practices actively inhibit anxiety escalation). Without structural psychological grounding, models generate conversations marked by superficial validation, premature client compliance, and ungrounded therapist interventions that fail to reflect the nuances of clinical psychotherapy.
This paper tackles this gap by leveraging Client Psychological Graphs (CPGs) from clinical psychology—directed networks where nodes signify atomic psychological processes and edges encode functional excitatory or inhibitory interactions. Because CPGs abstract away personal identifiers while preserving essential causal mechanisms, they provide an ideal, privacy-preserving substrate for generative modeling. Core idea: Ground synthetic counseling dialogue generation in process-based Client Psychological Graphs paired with empirically extracted counselor intervention strategies, steering LLM multi-turn interactions with clinical guidelines to synthesize rich, privacy-preserving, and psychologically coherent counseling dialogues.
Method¶
Overall Architecture¶
Graph2Counsel constitutes an end-to-end framework that translates structured clinical abstractions into naturalistic multi-turn counseling sessions and downstream evaluations. The pipeline comprises three central phases: first, CPG-grounded client profile generation and diversification, where privacy-scrubbed psychological graphs serve as the latent causal blueprint to synthesize comprehensive client intake forms across diverse demographic and contextual backgrounds; second, strategy-guided multi-turn dialogue generation, which integrates CPG topologies, client profiles, and counselor intervention strategies extracted from real therapy sessions under strict clinician-designed interaction constraints; and third, downstream model adaptation and multi-turn benchmarking, where open-source base models are fine-tuned via parameter-efficient techniques and evaluated on established benchmarks as well as a novel multi-turn interaction framework, Therapy-Eval.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Real Therapy Transcripts<br/>Clinical Data & Ethical Anonymization"] --> B["CPG Topology Extraction<br/>Nodes & Causal Edges"]
A --> C["Counselor Strategy Mining<br/>Intervention Techniques & Modalities"]
B --> D["CPG-Grounded Profile Inference<br/>Demographic & Narrative Diversification"]
D --> E["Strategy & Graph Guided Generation<br/>Clinical Constraints & Dialogue Synthesis"]
C --> E
B --> E
E --> F["Downstream QLoRA Fine-Tuning<br/>Llama3-8B Adaptation"]
F --> G["Therapy-Eval Benchmark<br/>Closed-Loop Agent Evaluation"]
Key Designs¶
1. Client Psychological Graph Abstraction: Modeling Functional Dynamics Without Privacy Leakage
To overcome the lack of psychological causality and the ethical perils of using raw transcripts, Graph2Counsel adopts Client Psychological Graphs as its primary structural representation. Grounded in Process-Based Therapy (PBT)—an evidence-based meta-framework integrating diverse therapeutic modalities—a CPG formalizes psychological processes (such as cognitive rumination, social anxiety, or behavioral avoidance) as nodes, connected by directed edges denoting activation (excitatory) or reduction (inhibitory) relations \(\langle \text{process}_i, \text{relation}, \text{process}_j \rangle\). Because the graph strips away personal identifiers while maintaining latent psychopathological interactions, it provides LLMs with an explicit causal map for reasoning about client reactions across turns while naturally safeguarding client confidentiality.
2. CPG-Grounded Profile Inference: Decoupling Psychological Structure from Demographic Context
To expand a modest set of seed clinical graphs into a diverse and scalable training distribution, the framework implements a structured profile inference and diversification pipeline. The same underlying cognitive network can manifest across vastly distinct real-world circumstances (for example, fear of judgment leading to workplace communication breakdown in a corporate setting versus marital conflict at home). Graph2Counsel prompts the generator to synthesize comprehensive client intake forms—spanning demographics, presenting problems, onset stressors, clinical history, and social support systems—conditioned on the CPG. Critically, the prompt enforces strict constraints forbidding the explicit mention of technical graph terminology or clinical jargon in client profiles, yielding 10 unique, realistic, and non-overlapping profiles per graph.
3. Counselor Strategy Mining: Grounding Interventions in Real-World Evidence-Based Modalities
Realistic therapy requires counselors to apply deliberate, evidence-based interventions rather than generic advice or interrogation-style questioning. Using an open-source LLM (Llama3.1-70B-Instruct), the framework extracts 257 distinct counselor intervention strategies alongside corresponding textual evidence directly from real therapy transcripts. These strategies span multiple established modalities, including Cognitive Behavioral Therapy (CBT), Dialectical Behavior Therapy (DBT), Acceptance and Commitment Therapy (ACT), and Motivational Interviewing (MI)—such as cognitive reframing, behavioral goal setting, and DEARMAN techniques. Supplying these extracted strategies as candidate options during generation enables the synthetic counselor to select clinically appropriate tools that match the unfolding therapeutic moment.
4. Strategy & Graph Guided Generation: Multi-Turn Dialogue Synthesis Under Clinical Constraints
To mitigate common pathological patterns observed in unconstrained LLM roleplay—such as clients agreeing too quickly, counselors firing rapid-fire questions, or dialogues concluding abruptly—the authors established rigorous behavioral guidelines in consultation with licensed clinical psychologists. Under the Guided Counseling (GC) setting, generation is subject to strict global constraints: client utterances must initially exhibit surface-level descriptions and respond to counselor reframing with hesitation, skepticism, or difficulty applying advice before reaching agreement; counselor turns must avoid ending every utterance with a question, must deliver psychoeducation through multi-sentence explanations (3–5 sentences), and must pace topic transitions across multiple exchanges. Comparing standard GC against Chain-of-Thought (GC+CoT) and multi-agent critique-refinement (GC+MA), GC with combined CPG and profile inputs achieved the optimal trade-off between computational cost, clinical rating, and natural conversational length (averaging 40.12 turns across 760 sessions).
Loss & Training¶
For downstream application, Llama3-8B-Instruct is fine-tuned using QLoRA across 14,597 generated \(\langle \text{dialogue\_history}, \text{counselor\_response} \rangle\) pairs. Training is conducted on 4 A100 80GB GPUs using DeepSpeed with LoRA rank \(r = 64\), scaling factor \(\alpha = 16\), learning rate \(1\times 10^{-5}\), dropout \(0.1\), batch size 4, and random seed 42 over 3 epochs.
Key Experimental Results¶
Main Results¶
Evaluation was conducted through an extensive blind review by four licensed American clinical psychologists, who ranked 100 semantically aligned dialogues across Graph2Counsel and three state-of-the-art synthetic baselines (CACTUS, MAGneT, SQPsychConv). Values represent average rank (lower is better; Safety denotes percentage of unsafe sessions). Downstream fine-tuned models were further evaluated on CounselBench-Eval and Therapy-Eval.
Expert Clinical Evaluation (Average Rank across 100 Matched Sessions):
| Dataset | Spec. (↓) | Compet. (↓) | Authent. (↓) | Safety (%) (↓) | Flow (↓) |
|---|---|---|---|---|---|
| CACTUS (Lee et al., 2024) | 2.41 | 2.39 | 2.22 | 3.0 | 2.00 |
| MAGneT (Mandal et al., 2026) | 3.97 | 3.98 | 3.99 | 1.0 | 3.99 |
| SQPsychConv (Vu et al., 2025) | 1.84 | 1.97 | 2.37 | 1.0 | 2.53 |
| Graph2Counsel (Ours) | 1.79 | 1.67 | 1.43 | 0.5 | 1.48 |
CounselBench-Eval Performance of Fine-Tuned Models:
| Model | Overall (↑, 1-5) | Empathy (↑, 1-5) | Specificity (↑, 1-5) | Medical Advice (%) (↓) | Factual Consistency (↑, 1-4) | Toxicity (↓, 1-5) |
|---|---|---|---|---|---|---|
| CAMEL (CACTUS) | 3.90 | 3.94 | 3.91 | 4.0 | 3.97 | 1.00 |
| Llama3-MAG | 4.12 | 4.09 | 4.12 | 8.0 | 3.98 | 1.00 |
| Llama3-SQP | 4.07 | 4.06 | 4.08 | 7.0 | 3.98 | 1.00 |
| Llama3-G2C (Ours) | 4.29 | 4.26 | 4.30 | 4.0 | 3.98 | 1.00 |
In Therapy-Eval multi-turn evaluation, Llama3-G2C achieved top ratings in specificity (3.81) and counselor competence (3.52) while sustaining an average of 30.57 engaging conversational turns, outperforming SQPsychConv's 22.56 turns.
Ablation Study¶
The authors ablated input representations (CPG, Profile, CPG+Profile) and prompting techniques (Base, GC, GC+CoT, GC+MA across 1–3 iterations) using automated LLM-as-a-judge scoring on the Cognitive Therapy Rating Scale (CTRS, 0–6) and Working Alliance Inventory (WAI, 1–7):
| Prompting Technique | Input Representation | CTRS General (U/I/C) | CTRS CBT (D/F/S) | WAI Task (↑) | WAI Goal (↑) | Avg. Turns |
|---|---|---|---|---|---|---|
| Base | CPG + Profile | 4.71 / 6.00 / 5.29 | 4.21 / 4.03 / 4.00 | 5.89 | 5.96 | 39.98 |
| GC (Ours Final) | CPG + Profile | 4.34 / 5.97 / 5.13 | 4.34 / 4.00 / 4.00 | 5.85 | 5.96 | 39.76 |
| GC + CoT | CPG + Profile | 4.42 / 6.00 / 5.39 | 4.32 / 4.00 / 4.00 | 5.96 | 6.19 | 30.72 |
| GC + MA (1 iter.) | CPG + Profile | 4.50 / 6.00 / 5.21 | 4.42 / 4.00 / 4.03 | 5.84 | 5.98 | 41.94 |
| GC + MA (3 iter.) | CPG + Profile | 4.53 / 6.00 / 5.00 | 4.21 / 4.00 / 4.00 | 5.75 | 5.85 | 43.26 |
Automated faithfulness verification demonstrated that 91% of target psychological processes from CPGs are manifested in the generated dialogues (CPG Faithfulness = 0.91), and profile contradiction was observed in only 1% of sessions (Profile Faithfulness = 0.99).
Key Findings¶
- Conditioning on explicit counselor strategies is crucial: compared to Base prompting, Guided Counseling consistently aligns counselor utterances with authentic clinical techniques and avoids premature topic shifting.
- While GC+CoT yields marginally higher automated CTRS/WAI scores, it substantially truncates dialogue length (reducing average turns from ~40 to ~30), which clinical evaluators disfavored compared to gradual session progression. GC+MA incurs prohibitive compute overhead without meaningful qualitative gains. Thus, GC with CPG+Profile represents the optimal production strategy.
- Expert agreement was remarkably high (Krippendorff’s \(\alpha = 0.70\) across qualitative dimensions, and 91% consensus on safety). Clinicians noted that Graph2Counsel dialogues featured significantly deeper, tailored interventions and avoided artificial, overly compliant client behavior.
Highlights & Insights¶
- Privacy-Preserving Causal Blueprints: Using CPGs abstracts therapeutic dynamics away from identifiable personal histories, enabling scalable diversification into hundreds of realistic client intake profiles while preserving clinical causality.
- Clinician-in-the-Loop Constraint Engineering: The integration of empirical clinical rules—such as requiring client resistance to initial reframing and disallowing sentence-ending interrogations—directly addresses the core failure modes of synthetic dialogue.
- Introduction of Therapy-Eval: Provides the community with a standardized multi-turn benchmark powered by 100 realistic depression character cards, bridging the gap between static multiple-choice QA benchmarks and multi-turn clinical reality.
Limitations & Future Work¶
- Constrained Seed Graph Diversity: The 76 CPGs were derived from only 6 patients at a single university psychology clinic, potentially inducing demographic and cultural homogeneity in the underlying causal structures. Multi-site clinical collaborations are needed to broaden graph variety.
- Static Graph Snapshots vs. Longitudinal Evolution: The current formulation models a single static snapshot per session. Real therapy involves longitudinal graph updates where cognitive patterns and interventions evolve across multi-week treatment trajectories.
- Evaluator Bias in Multi-Turn Benchmarks: Therapy-Eval utilizes off-the-shelf open-source client models and commercial judge LLMs without specialized clinical tuning, which may introduce residual evaluation biases.
Related Work & Insights¶
- vs CACTUS (Lee et al., 2024): CACTUS synthesizes large volumes of data from crowdsourced non-clinical personas, which clinicians found prone to repetitive, formulaic interactions; Graph2Counsel leverages real clinical CPGs and achieves superior authenticity and counselor competence.
- vs SQPsychConv (Vu et al., 2025): SQPsychConv conditions generation on discrete questionnaire symptom scores, which cannot capture directional, functional causal relations among thoughts and behaviors; Graph2Counsel directly models inhibitory and excitatory psychological dependencies.
- vs MDD-5k (Yin et al., 2025): MDD-5k samples triplets from tree structures solely for diagnostic background creation; Graph2Counsel uses the graph as an operational causal model governing full multi-turn therapeutic interventions.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ First framework to ground synthetic counseling dialogue generation in clinical psychological graphs and real-world intervention strategies.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous blind expert evaluation by four licensed clinicians, comprehensive automated CTRS/WAI ablation, and thorough downstream benchmark verification.
- Writing Quality: ⭐⭐⭐⭐⭐ Clear, methodically sound, with exceptionally rich appendices detailing clinical prompts, evaluation protocols, and error analyses.
- Value: ⭐⭐⭐⭐⭐ Provides a viable, privacy-compliant blueprint for generating safe, clinically grounded training data for healthcare LLMs.