Source-linked AI summary
TheraMind: A Strategic and Adaptive Agent for Longitudinal Psychological Counseling
He Hu, Chiyuan Ma, Qianning Wang, Lin Liu, Yucheng Zhou, Laizhong Cui, Fei Ma, Qi Tian
TL;DR
LLM-based counseling offers scalable support, but existing approaches often lack emotional understanding, adaptive strategies, and long-term memory. TheraMind uses dual loops for turn-level dialogue and cross-session planning; evaluations report strong results, especially on multi-session measures.
Problem
Existing LLM counseling approaches often lack emotional understanding, adaptive strategies, and long-term memory.
Source: §1 · p. 1Method
TheraMind combines turn-level state perception and response-strategy selection with cross-session efficacy evaluation and therapy adjustment.
Source: §1 · p. 2, §3 · p. 3, §3.1 · p. 4Results
Takeaways & Limitations
The dual-loop architecture supports longitudinal counseling by separating in-session dialogue management from cross-session therapeutic planning.
Source: §5 · pp. 8–9
Abstract
from arXiv · showhide
The shortage of mental health professionals has driven the web to become a primary avenue for accessible psychological support. While Large Language Models (LLMs) offer promise for scalable web-based counseling, existing approaches often lack emotional understanding, adaptive strategies, and long-term memory. These limitations pose risks to digital well-being, as disjointed interactions can fail to support vulnerable users effectively. To address these gaps, we introduce TheraMind, a strategic and adaptive agent designed for trustworthy online longitudinal counseling. The cornerstone of TheraMind is a novel dual-loop architecture that decouples the complex counseling process into an Intra-Session Loop for tactical dialogue management and a Cross-Session Loop for strategic therapeutic planning. The Intra-Session Loop perceives the patient's emotional state to dynamically select response strategies while leveraging cross-session memory to ensure continuity. Crucially, the Cross-Session Loop empowers the agent with long-term adaptability by evaluating the efficacy of the applied therapy after each session and adjusting the method for subsequent interactions. We validate our approach in a high-fidelity simulation environment grounded in real clinical cases. Extensive evaluations show that TheraMind outperforms other methods, especially on multi-session metrics like Coherence, Flexibility, and Therapeutic Attunement, validating the effectiveness of its dual-loop design in emulating strategic, adaptive, and longitudinal therapeutic behavior. The code is publicly available at https://github.com/Emo-gml/TheraMind.
1 Introduction
TheraMind addresses the limits of LLM counseling systems that lack cross-session memory and flexible treatment strategies with a dual-loop agent for longitudinal counseling. Experiments and human evaluations report strong performance, particularly on multi-session measures.
1 Introduction
- Existing LLM counseling systems often focus on single sessions, lack cross-session memory, and use fixed therapeutic modalities, limiting continuity and adaptability.The introduction frames these shortcomings as clinical amnesia and strategic rigidity.Source: §1 · pp. 1–2
- TheraMind uses an Intra-Session Loop for tactical dialogue and a Cross-Session Loop for strategic planning and therapy adjustment across sessions.It perceives emotional state to select response strategies and evaluates efficacy after each session.Source: §1 · p. 2
- TheraMind significantly outperforms baselines in a high-fidelity simulation, with gains especially pronounced in multi-session metrics; ablations and human evaluations also support its performance.The simulation is grounded in real clinical cases.Source: §1 · p. 2
- Its contributions combine turn-level patient-state perception and strategy selection with treatment-phase awareness and adaptive therapy selection across sessions.The design aims to support deliberative interventions and address static single-therapy approaches.Source: §1 · p. 2
2 Related Work
Prior counseling agents and LLM-based systems pursue empathy, therapeutic techniques, and longer-term support, while gaps in emotional understanding and adaptation remain. Related work also surveys autonomous-agent architectures and their mental-health applications.
2 Related Work
- Across these approaches, emotional understanding, adaptive strategies, and long-term memory remain gaps in LLM-based counseling.Source: §2 · p. 2
2.1 Counseling Dialogue Systems
- Counseling dialogue systems span pattern matching, emotion and dialogue-act methods, structured therapies, and fine-tuned LLMs, including ChatCounselor [24], PsyLLM [16], and PsyDial.PsyDial builds longitudinal datasets, while PsyDT [45] uses dynamic one-shot learning and GPT-4–guided synthesis.Source: §2.1 · p. 2
- Supervised fine-tuning can struggle to unify capabilities, while reinforcement learning is sensitive to reward design, leaving challenges in adapting to therapeutic complexity [11] [21] [39].Source: §2.1 · p. 3
2.2 LLM-driven Autonomous Agents
- Autonomous-agent research combines reasoning, planning, memory, and tool use, with mental-health applications including safety, therapeutic role-play, and structured interviews.Examples include AnnaAgent [40] for multi-session memory, EmoAgent [26] for safety, and AutoCBT [46] for CBT-based multi-agent counseling.Source: §2.2 · p. 3
3 Methodology
TheraMind addresses counseling agents’ limited ability to sustain strategic, adaptive therapy across sessions with a dual-loop architecture. Its Intra-Session Loop manages turn-level dialogue while the Cross-Session Loop plans and adapts the treatment trajectory.
3 Methodology
- The dual-loop architecture separates tactical turn-by-turn dialogue management from strategic planning and adaptation across multiple sessions.This hierarchy is designed to maintain immediate conversational coherence alongside long-term therapeutic objectives.Source: §3 · p. 3
3.1 The Intra-Session Loop: Dynamic Dialogue Management
- The Intra-Session Loop processes each conversational turn through a multistage perception-to-action process.It is the agent’s real-time interaction engine.Source: §3.1 · p. 4
3.1.1 Multi-Faceted Patient State Perception.
- The Reaction Classifier maps a patient utterance to a structured state comprising primary emotion, its intensity, and conversational attitude.The attitude is represented as Cooperative or Resistant, and emotion intensity ranges from 0 to 1.Source: §3.1.1 · p. 4
3.1.2 Memory-Augmented Contextualization.
- Memory Recall uses the current utterance and prior-session history to produce a relevant memory summary, or null when no pertinent history is found.The system uses this dynamically curated context to inform responses with relevant history.Source: §3.1.2 · p. 4
3.1.3 Clinically-Grounded Response Generation.
- Response generation combines the perceived patient state, a state-dependent tactical strategy, the current therapeutic phase, and the long-term therapy method.The response is generated by an LLM conditioned on these factors.Source: §3.1.3 · p. 4
3.2 The Cross-Session Loop: Adaptive Therapeutic Planning
TheraMind evaluates the therapy used in each session and uses that feedback to select a method for the next session. Selection considers efficacy alongside dialogue history, allowing the current method to be maintained or changed.
3.2 The Cross-Session Loop: Adaptive Therapeutic Planning
- The cross-session loop is presented as enabling reflection and adaptation of the treatment strategy over time.The paper describes this as the agent’s departure from prior work.Source: §3.2 · p. 4
3.2.1 Post-Session Efficacy Evaluation.
- After each session, a Therapy Evaluation module assesses the efficacy of the method used and produces a quantitative score.The assessment uses the session’s complete dialogue history and the employed therapeutic method.Source: §3.2.1 · p. 4
3.2.2 Adaptive Therapy Selection.
- TheraMind’s LLM-based Therapy Selection module interprets the efficacy score in the context of session history to make a qualitative clinical judgment.It can maintain the current method when progress is sufficient or select a more suitable method based on outcomes and shortcomings.Source: §3.2.2 · p. 4
3.3 High-Fidelity Longitudinal Simulation Environment
The paper evaluates longitudinal counseling with a dynamic, high-fidelity simulation grounded in real clinical cases. Structured session guides aim to balance clinical realism with reproducibility.
3.3 High-Fidelity Longitudinal Simulation Environment
- The proposed evaluation paradigm uses dynamic, high-fidelity simulation to assess longitudinal counseling capabilities that static, single-turn datasets are ill-suited to evaluate.The targeted capabilities include long-term memory, strategic adaptation, and therapeutic alliance building.Source: §3.3 · p. 5
- The simulation samples 100 anonymized CPsyCounR cases, with 10 cases selected from each of the corpus’s 10 clinical categories.Evaluations use a six-session dialogue arc.Source: §3.3 · p. 5
- High-level session guides give the patient simulator coherent goals without rigidly scripting replies, supporting dynamic emotional responses and a testbed balancing clinical realism with reproducibility.The guides are generated alongside a structured patient profile for a six-session arc.Source: §3.3 · p. 5
4 Experiments
TheraMind was evaluated in simulated longitudinal counseling using clinical cases, with results across automated metrics, human judgments, and analyses of its behavior. It outperformed baselines, with especially strong multi-session results, while component ablations and case analysis examined how its design operates.
4.2 Evaluation Dataset and Metrics
- Evaluation used 3,134 clinical counseling reports and a six-session simulation of 100 sampled cases; trained annotators rated single- and multi-session metrics on a 0–3 scale.Longitudinal measures include continuity, strategy flexibility, empathy, and therapeutic attunement.Source: §4.2.1 · p. 5, §4.2.2 · p. 5, §4.3 · p. 6
4.4 Main Results
- TheraMind significantly outperformed all baselines across single- and multi-session metrics, leading in Coherence (2.860), Empathy (2.980), Therapeutic Attunement (2.890), and Flexibility (2.290).Against DeepSeek-V3, its multi-session average rose from 2.330 to 2.755, an 18.2% relative improvement.Source: Table 2 · p. 6, §4.4 · p. 6
4.5 Ablation Study
- The full model beat every ablated variant; removing Memory Mechanism yielded a 75.00% win rate, compared with 72.73% for Stage Adaptation, 68.42% for Response Strategy, and 66.67% for Therapy Selection.Figure 3 compares the full model pairwise against variants missing each component.Source: §4.5 · p. 6, Figure 3 · p. 6
4.6 Analysis Across Psychological Issues
- Across ten issue categories, TheraMind led in single-session scores and maintained consistently high multi-session performance as competing models’ gains remained inconsistent and marginal.The performance gap widened in multi-session evaluation.Source: §4.6 · pp. 6–7, Figure 4 · p. 7
4.7 Human Evaluation
- Human evaluations found substantial agreement (Cohen’s κ 0.676) and preference win rates of 75% against PsyDTLLM, 70% against DeepSeek-V3, and 65% against Interactive Agents.Emotion Recognition and Memory Invoke agreement were 95.59% and 94.12%, respectively; Response Strategy agreement was 85.29%.Source: §4.7 · p. 7, Figure 5 · p. 7
4.8 In-depth Analysis of TheraMind’s Behavior
- TheraMind’s intervention repertoire combined Reflection of Feelings (32.54%) with perspective-taking (16.93%) and action-invitation (15.83%) strategies, while sessions progressed from Engagement through Exploration to Integration.Cooperative attitudes correlated with higher emotional intensity; resistant attitudes corresponded to lower, more dispersed intensity.Source: §4.8 · p. 8
4.9 Case Study
- In a Session 3 case, the agent paired the patient’s sadness and negative attitude with a retrieved Session 2 “rib” metaphor about the mother’s expectations.The response connected the current distress to that prior insight to maintain continuity.Source: §4.9 · p. 8, Figure 7 · p. 8
5 Conclusion
TheraMind uses a dual-loop architecture to address clinical amnesia and strategic rigidity. Experiments and human evaluations support its state-of-the-art performance and long-term coherence and flexibility.
5 Conclusion
- TheraMind’s dual-loop architecture addresses prior work’s clinical amnesia and strategic rigidity.It separates tactical in-session dialogue management from strategic cross-session therapeutic planning.Source: §5 · pp. 8–9
- Experiments and human evaluations confirmed TheraMind’s state-of-the-art performance.The conclusion highlights superior long-term coherence and flexibility.Source: §5 · p. 9
- The architecture separates in-session dialogue management from cross-session therapeutic planning, supporting longitudinal counseling.Source: §5 · p. 9
A Human Evaluation of Pairwise Response Selection
Human evaluators compared TheraMind with three systems across counseling cases and sessions, reporting better counseling effects for TheraMind and substantial evaluator consistency.
A Human Evaluation of Pairwise Response Selection
- TheraMind demonstrated better counseling effects than the comparison systems across the evaluated psychological counseling scenarios.The evaluation counted wins across 20 samples and 120 sessions, comparing it pairwise with PsyDTLLM, DeepSeek-V3, and Interactive Agents.Source: A Human Evaluation of Pairwise… · p. 11
- The evaluation randomly sampled two cases from each of ten categories and used trained annotators to compare responses in three pairwise groups.The comparisons were conducted on a self-designed platform (Figure 8).Source: A Human Evaluation of Pairwise… · p. 11, Figure 8 · p. 11
- Cohen’s κ was 0.697, which the authors interpret as indicating high objectivity and impartiality in the human evaluation.The paper describes this as generally reliable manual evaluation.Source: A Human Evaluation of Pairwise… · p. 11
B Human Agreement Analysis
Annotators assessed three decision-making modules using dialogue context from current and previous sessions. Reported agreement rates were highest for Emotion Recognition and Memory Invoke, with substantial inter-rater agreement.
B Human Agreement Analysis
- The assessment covered 68 dialogue rounds, using one randomly selected case from each of ten categories and one session per case.Two psychology-trained annotators received each target session and previous sessions for context; Figure 9 shows the evaluation platform.Source: B Human Agreement Analysis · p. 11, Figure 9 · p. 11
- Emotion Recognition reached 95.59% agreement, Memory Invoke 94.12%, and Response Strategy 85.29%.The authors describe these results as evidence of fidelity to human clinical reasoning and clinically coherent decisions.Source: B Human Agreement Analysis · p. 11
- Cohen’s κ was 0.67, within the range the paper identifies as substantial agreement.Source: B Human Agreement Analysis · p. 11
C Ethical Considerations
The authors report privacy safeguards for their dataset, fair compensation for trained annotators, and limited human expert assessment as a response to evaluation bias concerns.
C Ethical Considerations
- The study used anonymized, professionally rewritten Chinese counseling reports from the CPsyCounR dataset under academic-research ethical guidelines.Safeguards were used to prevent re-identification or misuse.Source: C Ethical Considerations · p. 11
- Trained psychology annotators were paid above local minimum wage, in line with standard academic research practices.Source: C Ethical Considerations · p. 11
- Because long dialogue histories make multi-session evaluation challenging and LLM-based methods may introduce bias, the study used limited human expert assessments.Source: C Ethical Considerations · p. 11
D Expert Evaluation
A same-case, same-session comparison contrasts TheraMind with PsyDTLLM. TheraMind is described as better supporting emotional processing, treatment continuity, and therapeutic pacing.
D Expert Evaluation
- In the same case and session, TheraMind better supports emotional processing and treatment continuity than PsyDTLLM.The comparison is described as a randomly selected case analysis.Source: D Expert Evaluation · pp. 11–12, Figure 10 · p. 12
- PsyDTLLM overlooks subtle emotional signals, moves to techniques before sufficient emotional exploration, and sometimes underuses historical dialogue.The passage associates these limitations with insufficient emotional-processing depth and poor treatment continuity.Source: D Expert Evaluation · p. 11
- TheraMind guides patients toward root causes of negative emotions while adjusting therapeutic pace and managing session boundaries.Its patient-centered approach makes emotional understanding the entry point for intervention.Source: D Expert Evaluation · p. 12