Organizations: Dongwu Business School, Soochow University, Suzhou, China · School of Computer Science and Technology, Soochow University, Suzhou, China · College of Computing and Data Science, Nanyang Technological University, Singapore
Mental health disorders affect hundreds of millions of people around the world, yet access to professional counseling remains severely limited. AI-powered dialogue systems offer a scalable alternative, but existing models face two fundamental challenges. First, they lack the bidirectional understanding needed to capture the layered nature of emotional expression, particularly in cases of progressive disclosure, where clients often present symptoms at the surface-level while concealing deeper trauma. Autoregressive (AR) models process information sequentially and cannot revise early interpretations when new evidence emerges later in the conversation. Second, they fail to effectively incorporate the relational knowledge that underlies clinical reasoning. In this paper, we propose BiGraph-Diffuse, the first large-scale diffusion language model tailored for the counseling domain. We further introduce BiGraph-RAG, a relation-free graph-structured retrieval strategy that relies only on lightweight entity extraction and semantic linking. This design preserves inferential pathways from observable symptoms to potential underlying causes, while incurring zero LLM token cost during indexing. Importantly, these two modules are not merely combined but mutually reinforcing. The diffusion model provides a holistic bidirectional context, enabling the system to defer premature judgments during progressive disclosure. Meanwhile, graph-based retrieval captures the structured interconnections of clinical knowledge. Extensive experiments demonstrate the effectiveness of BiGraph-Diffuse, and we further provide a solid theoretical analysis to support its design.
Figures & tables
Fig. 1: Motivation of BiGraph-Diffuse. Left: Autoregressive models fail to suspend judgment on early defensive disclosures (e.g., insomnia) and cannot revise shallow advice when deep trauma (e.g., bullying) is later revealed. Right: Traditional RAG retrieves isolated chunks and breaks the inferential chain between surface symptoms and underlying causes.
Fig. 2: Overview of the BiGraph-Diffuse framework. The system consists of three stages: (1) Offline indexing constructs a heterogeneous Tri-Graph from authoritative psychology corpora; (2) During inference, BiGraph-RAG performs two-stage propagation over the graph to retrieve relevant knowledge; (3) The fine-tuned BiGraph-Diffuse model generates a thought process and final response conditioned on both the retrieved knowledge and dialogue history.
Model
OpenR1-Psy (English)
CPsyCounE (Chinese)
E&I
S&A
A&P
S&B
Norm. AVG
E&I
S&A
A&P
S&B
Norm. AVG
Autoregressive Baselines (No Retrieval)
CPsyCounX [ 8 ]
0.012
0.122
0.061
0.357
0.138
0.438
0.528
0.462
0.818
0.562
ChatCounselor [ 10 ]
0.020
0.207
0.110
0.470
0.202
—
—
—
—
—
MeChat [ 12 ]
0.024
0.213
0.139
0.799
0.294
0.224
0.312
0.253
0.591
0.345
PsyDTLLM [ 33 ]
0.073
0.285
0.229
0.873
0.365
0.193
0.283
0.242
0.621
0.335
TABLE I: Performance comparison on OpenR1-Psy and CPsyCounE test sets.
Fig. 3: Impact of LoRA fine-tuning and graph retrieval on normalized average score.
Model
OpenR1-Psy
CPsyCounE
Norm. AVG
Norm. AVG
PsyLLM-8B + BiGraph-RAG
0.621 ± 0.013
—
CPsyCounX + BiGraph-RAG
—
0.611 ± 0.016
PsyLLM-8B + BiGraph-RAG + Suspend
0.632 ± 0.011
—
CPsyCounX + BiGraph-RAG + Suspend
—
0.622 ± 0.012
GPT-4o + BiGraph-RAG + Suspend
0.585 ± 0.013
0.643 ± 0.013
TABLE II: AR + BiGraph-RAG + suspend-judgment prompting vs. BiGraph-Diffuse. All scores are normalized averages over 3 independent runs ( ± std).
Retrieval Backbone
OpenR1-Psy
Δ
CPsyCounE
Δ
(English)
(vs. w/o)
(Chinese)
(vs. w/o)
w/o Retrieval
0.785 ± 0.010
—
0.834 ± 0.012
—
+ Dense RAG
0.798 ± 0.012
+0.013
0.851 ± 0.013
+0.017
+ LinearRAG
0.806 ± 0.011
+0.021
0.859 ± 0.012
+0.025
+ BiGraph-RAG
0.824 ± 0.009
+0.039
0.883 ± 0.014
+0.049
TABLE III: Comparison of retrieval backbones with the same BiGraph-Diffuse generator. Δ indicates absolute gain over the no-retrieval baseline.
Fig. 4: Layperson average ranking (lower is better).
TABLE V: Quantitative analysis of retrieval quality on psychology corpus.
Metric
Simple Fact
Emp./Support
Prog. Disc.
Avg.
Relevance
0.880
0.830
0.790
0.833
Hit Rate@1
0.850
0.780
0.720
0.783
Evidence Chain Compl.
—
—
0.740
0.740
TABLE VI: Retrieval accuracy across query types.
Fig. 5: Absolute gain from BiGraph-RAG averaged across OpenR1-Psy and CPsyCounE.
Model
Integration
Revision
Sensitivity
Temporal
Norm. AVG
GPT-4o [ 35 ]
0.534
0.487
0.568
0.751
0.561
DeepSeek-V3 [ 34 ]
0.516
0.479
0.589
0.776
0.563
BiGraph-Diffuse (w/o retrieval)
0.645
0.603
0.698
0.874
0.681
BiGraph-Diffuse
0.811
0.763
0.868
0.949
0.833
TABLE VII: Performance on progressive disclosure subset (normalized 0–1 scores).
Model
Gemini
GPT-4.1 Mini
Δ
(w/ thought)
(w/o thought)
GPT-4o + BiGraph-RAG + Suspend
0.585
0.571
− 0.014
DeepSeek-V3 + BiGraph-RAG + Suspend
0.573
0.558
− 0.015
BiGraph-Diffuse (w/o retrieval)
0.785
0.778
− 0.007
BiGraph-Diffuse
0.824
0.817
− 0.007
TABLE VIII: Multi-judge comparison with and without internal thought trace. Gemini scores include thought trace; GPT-4.1 Mini scores are based on final response only.
Fig. 6: Layer-wise probing accuracy for predicting future deep disclosure from first-turn text.
Effective mental health counseling is a complex, theory-driven process requiring the simultaneous integration of psychological frameworks, real-time distress signals, and strategic intervention planning. This level of clinical reasoning is critical for safety and therapeutic effectiveness but is often missing in general-purpose Large Language Models (LLMs). We introduce SAGE (Strategy-Aware Graph-Enhanced), a novel framework designed to bridge the gap between structured clinical knowledge and generative AI. SAGE constructs a heterogeneous graph that unifies conversational dynamics with a psychologically grounded layer, explicitly anchoring interactions in a theory-driven lexicon. Our architecture first employs a Next Strategy Classifier to identify the optimal therapeutic intervention. Subsequently, a Graph-Aware Attention mechanism projects graph-derived structural signals into soft prompts, conditioning the LLM to generate responses that maintain clinical depth. Validated through both automated metrics and expert human evaluation, SAGE outperforms baselines in strategy prediction and recommended response quality. By providing actionable intervention recommendations, SAGE serves as a cutting-edge decision-support tool designed to augment human expertise in high-stakes crisis counseling.
Eliya Naomi Aharon, Meytal Grimland, Avi Segal +4
Ben-Gurion University of the Negev Beer Sheva, Israel · University of Haifa Haifa, Israel · Sahar Tel Aviv, Israel
Rising demand for mental health support has increased interest in using Large Language Models (LLMs) for counseling, but adapting them to this safety-critical domain is hindered by limited real-world data due to privacy constraints. Synthetic datasets provide a promising alternative, but existing approaches often rely on unstructured or semi-structured text inputs and overlook structural dependencies between a client's cognitive, emotional, and behavioral states, leading to psychologically inconsistent and less realistic interactions. We introduce Graph2Counsel, a framework for generating synthetic counseling sessions grounded in Client Psychological Graphs (CPGs) that encode relationships among clients' thoughts, emotions, and behaviors. Graph2Counsel uses a structured prompting pipeline guided by counselor strategies and CPG, and explores prompting strategies including CoT and Multi-Agent Feedback. It produces 760 sessions from 76 CPGs across diverse client profiles. In expert evaluation, our dataset outperforms prior datasets on specificity, counselor competence, authenticity, conversational flow, and safety (Krippendorff's α = 0.70). Fine-tuning an open-source model on this dataset improves performance on several metrics on CounselingBench and CounselBench, while matching baselines on others. We further introduce Therapy-Eval, a multi-turn evaluation framework, and demonstrate the effectiveness of our fine-tuned model in realistic therapeutic conversations. We make our code, data and fine-tuned model public.
Aishik Mandal, Hiba Arnaout, Clarissa W. Ong +5
UKP Lab, Department of Computer Science and Hessian Center for AI (hessian.AI), Technische Universität Darmstadt · Zuse School · National Research Center for Applied Cybersecurity ATHENE +4
Large language models (LLMs) show promise in generating supportive responses for mental health and counseling applications. However, their responses often lack cultural sensitivity, contextual grounding, and clinically appropriate guidance. This work addresses the gap of how to systematically incorporate domain-specific, clinically validated knowledge into LLMs to improve counseling quality. We utilize and compare two approaches, retrieval-augmented generation (RAG) and a knowledge graph (KG)-based method, designed to support para-counselors. Our KG is constructed manually and clinically validated, capturing causal relationships between stressors, interventions, and outcomes, with contributions from multidisciplinary people. We evaluated multiple LLMs in both settings using BERTScore F1 and SBERT cosine similarity, as well as human evaluation across five metrics, which is designed to directly measure the effectiveness of counseling beyond similarity at the surface level. The results show that KG-based approaches consistently improve contextual relevance, clinical appropriateness, and practical usability compared to RAG alone, demonstrating that structured, expert-validated knowledge plays a critical role in addressing LLMs limitations in counseling tasks.
Md Arid Hasan, Azhagu Meena SP, Aditya Khan +6
University of Toronto, Canada · mPower Social Enterprises Ltd, Bangladesh · National Institute of Mental Health, Bangladesh +2