PaperlayerView exact revision on arXiv

Source-linked AI summary

Psychological Counseling Cannot Be Achieved Overnight: Automated Psychological Counseling Through Multi-Session Conversations

Junzhe Wang, Bichen Wang, Xing Fu, Yixin Sun, Yanyan Zhao, Bing Qin

arXiv:2506.06626v1cs.CL

TL;DR

Automated psychological counseling research has focused on single sessions, although counseling often involves evolving client states and goals across sessions. The paper introduces a CBT-based dataset built from public case reports and a model for multi-session counseling, reporting ability across multiple psychological metrics.

  • Problem

    Automated counseling research has focused on single sessions, while counseling often spans multiple sessions in which client states and counselor goals evolve.

    Source: §1 · p. 1

  • Method

    The paper constructs a CBT-based multi-session counseling dataset from public case reports and trains a model to track client progress and adapt counseling direction.

    Source: §1 · p. 2

  • Results

    MusPsy-Model demonstrates ability across basic indicators, the Working Alliance Inventory, and the Positive and Negative Affect Schedule.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The dataset offers a resource for developing LLM-based systems that better approximate the complexities of real-world counseling.

    Source: §7 · p. 8

    The authors cite cost constraints and challenges inherent to the field, and say they will not directly promote the system as a commercial service to avoid harming vulnerable populations.

    Source: Limitations · p. 9

Abstract

from arXiv · show

In recent years, Large Language Models (LLMs) have made significant progress in automated psychological counseling. However, current research focuses on single-session counseling, which doesn't represent real-world scenarios. In practice, psychological counseling is a process, not a one-time event, requiring sustained, multi-session engagement to progressively address clients' issues. To overcome this limitation, we introduce a dataset for Multi-Session Psychological Counseling Conversation Dataset (MusPsy-Dataset). Our MusPsy-Dataset is constructed using real client profiles from publicly available psychological case reports. It captures the dynamic arc of counseling, encompassing multiple progressive counseling conversations from the same client across different sessions. Leveraging our dataset, we also developed our MusPsy-Model, which aims to track client progress and adapt its counseling direction over time. Experiments show that our model performs better than baseline models across multiple sessions.

1 Introduction

The introduction argues that automated counseling should model the evolving goals and client states of multi-session counseling, rather than isolated sessions. It presents a CBT-based dataset and model designed for this setting, with the model evaluated across multiple psychological metrics.

1 Introduction

  • The paper identifies growing demand for mental health support alongside scarce accessible services as motivation for automated counseling research.Source: §1 · p. 1
  • Counseling often spans sessions, with counselors recalibrating goals as clients’ mental states change; single-session research does not capture this progression.The paper frames sustained, coherent multi-session conversations as a need for automated counseling models.Source: §1 · p. 1
  • The paper integrates a multi-session paradigm to model counseling progression and adjust counseling goals as client states change.It proposes that this approach can cumulatively build rapport, address clients’ problems, and support long-term psychological improvement.Source: §1 · pp. 1–2
  • MusPsy-Model is trained on MusPsy-Dataset and evaluated across basic indicators, WAI, and PANAS.The introduction reports that the model demonstrates ability across multiple psychological metrics.Source: §1 · p. 2
  • MusPsy-Dataset uses CBT-based, multi-session conversations generated from client profiles and counseling goals in public case reports.The construction process first generates short seed conversations, then elaborates them into complete, consistent, goal-driven conversations across sessions, with session-level memories extracted to support continuity.Source: §1 · p. 2

2 Related Work

The related work describes CBT as a structured process that unfolds over multiple sessions, then surveys automated counseling systems and distinguishes this work by its multi-session focus.

2.1 Cognitive Behavioral Therapy

  • CBT identifies and challenges unhelpful thoughts and beliefs over a structured, multi-session process to support more positive beliefs over time.The section describes counselors first helping clients recognize unhelpful thoughts, then guiding them in challenging and correcting distortions.Source: §2.1 · p. 2

2.2 Automated Psychological Counseling

  • Prior work includes SMILE and CBT-LLM (Qiu et al., 2024), CPsyCoun (Zhang et al., 2024), Healme (Xiao et al., 2024), and single-turn CBT processes in CACTUS (Lee et al., 2024); this work adds multi-session dynamics.The section also notes data scarcity tied to privacy concerns and protection of vulnerable groups.Source: §2.2 · p. 2

3 MusPsy-Dataset: A Multi-Session Psychological Counseling Conversation Dataset

MusPsy-Dataset is constructed from longitudinal client profiles and goals drawn from case reports, then generated through staged counseling and memory processes. Its design addresses continuity and context limits in multi-session generation.

3 MusPsy-Dataset: A Multi-Session Psychological Counseling Conversation Dataset

  • The dataset construction has four stages: collecting profiles and goals, building seed conversations, generating counseling, and generating memory.These stages together produce the MusPsy-Dataset.Source: §3 · p. 3, Figure 2 · p. 4

3.1 Client Profiles & Counseling Goals Collection

  • Profiles from published case reports capture static traits, session-varying states, and planned goals, providing authenticity and continuity without full transcripts.The reports document client states and goals achieved per session, but exclude full session transcripts for privacy reasons.Source: §3.1 · p. 3

3.2 Seed Conversation Construction

  • Concurrent seed conversations (3–4 turns; about 100–200 tokens) guide full multi-session generation, addressing one-pass depth limits and weak continuity from separate sessions.The paper describes seeds as concise conversations generated simultaneously to establish coherence across sessions.Source: §3.2 · p. 3

3.3 Counseling Generation

  • Full sessions are expanded from seed conversations with preceding seeds and CBT examples, preserving temporal and emotional progression while guiding counselor-like tone and technique.The prompts also use real counseling fragments from reports as examples for relevant client-profile stages.Source: §3.3 · pp. 3–4

3.4 Memory Generation

  • Session-end memory condenses profiles, counseling notes, and a series summary, helping later sessions retain key information without storing full transcripts.The memory module addresses context-window and consistency challenges in long-term counseling.Source: §3.4 · pp. 4–5

4 Data Statistics and Evaluation

The evaluation compares MusPsy-Dataset with prior counseling datasets and examines both single-session quality and session-by-session working alliance. The authors report broad quality advantages and a general upward alliance trend, with a sixth-session exception in goal agreement.

4.1 Data Statistics

  • 1,400 clients contribute multi-turn CBT counseling, averaging 6.17 sessions per client, 28.55 turns per session, and 26.99 words per turn.The authors describe this scale as sufficient to support construction of their multi-session counseling model.Source: §4.1 · p. 5, Table 1 · p. 5

4.2 Competing Datasets

  • Prior counseling datasets include SMILE, CACTUS, SimPsyDial, and CPsyCoun, with their reported data quality and model performance considered in the evaluation framework.Source: §4.2 · p. 5

4.3 Dataset Quality Evaluations

  • Dataset quality is assessed on four basic metrics and three WAI dimensions, using ratings from two psychology experts and GPT-4o on a 1–5 scale.The basic metrics are helpfulness, empathy, guidance, and coherence; WAI dimensions are Goal Agreement, Task Agreement, and Emotional Bond.Source: §4.3 · p. 5, Table 2 · p. 6

4.3.1 Single-Session Evaluation

  • MusPsy-Dataset outperforms existing counseling datasets across nearly all dimensions in the single-session quality evaluation.Table 2 covers four basic metrics and three WAI dimensions, assessed by GPT-4o and manual evaluation.Source: Table 2 · p. 6, §4.3.1 · p. 6
  • Strong Emotional Bond, Task Agreement, and Goal Agreement are attributed to gradual counseling progression and session-by-session goal setting, though WAI metrics still have room to improve.The authors link the dataset’s quality and applicability to simulating counseling as a process rather than compressing it into one session.Source: §4.3.1 · p. 6

4.3.2 Multi-Session Evaluation

  • Across up to six sessions, WAI scores generally rise, but goal agreement dips slightly in session six as continuing clients address deeper, harder-to-accept goals.The authors note that some clients finish counseling around five sessions, while those continuing to six may work on core beliefs.Source: Figure 4 · p. 6, §4.3.2 · p. 6

5 Task Setting

The task setting divides multi-session counseling into three linked capabilities: extracting session memory, planning the next session’s goals, and generating counseling informed by that memory and those goals.

5 Task Setting

  • Together, the three tasks mirror a human counselor’s workflow for managing counseling progress and generating coherent sessions.Source: §5 · p. 7, Figure 5 · p. 7
  • Memory Extraction updates the model’s memory with key information from the current session.Source: §5 · p. 7
  • Goal Planning uses the model’s memory to plan counseling goals for the next session.Source: §5 · p. 7
  • Counseling Generation uses past memory and planned goals to guide the next counseling session.Source: §5 · p. 7

6 Experiment

The experiments evaluate memory extraction, goal planning, and multi-session counseling generation. The authors report that MusPsy-Model improves client emotional state and counseling performance over multiple sessions.

6.1 Experiment Setup

  • Experiments use Meta-Llama-3-8B-Instruct, with other datasets trained on the same model for comparability.Prompts and hyperparameters follow those specified in the original papers.Source: §6.1 · p. 7

6.2 Evaluations and Results of Task 1&2

  • Memory Extraction and Goal Planning can update memory from session content and predict relevant goals for the following session.The authors describe these capabilities as a foundation for progressive multi-session counseling.Source: §6.2 · p. 7

6.3 Evaluations and Results of Task 3

  • For Counseling Generation, the model receives previous-conversation memory and the current session’s goal as input.The evaluation uses multi-session simulations with 100 test client profiles, with an LLM acting as the client.Source: §6.3 · p. 7

6.3.1 Evaluation Results

  • Across multiple sessions, MusPsy-Model improves client emotional state and counseling performance, while most other models show little benefit.After two sessions, other models largely repeat prior counseling; after six sessions, MusPsy-Model significantly alleviates negative emotions.Source: Figure 6 · p. 8, §6.3.1 · p. 8

7 Conclusion

The paper presents MusPsy-Dataset and MusPsy-Model as an exploratory step toward automated counseling that represents multi-session progress. Its initial results are promising, and the authors identify broader therapeutic diversity as future work.

7 Conclusion

  • MusPsy-Dataset and explicit modeling of client progress offer a resource for developing LLM-based systems that better approximate real-world counseling complexities.Source: §7 · p. 8
  • Initial MusPsy-Model results are promising, with the model tracking client state and adapting counseling goals over time.Source: §7 · p. 8
  • Future work should expand the diversity of therapeutic approaches represented in the dataset.Source: §7 · p. 9

Limitations

The authors identify limitations in therapeutic scope, client coverage, realism, evaluation, privacy, and readiness for deployment.

Limitations

  • The dataset and model use CBT, which may not suit all clients, and the model lacks flexible, multi-modal counseling strategies.Source: Limitations · p. 9
  • Clients resistant to treatment receive insufficient attention in both dataset synthesis and model evaluation.Source: Limitations · p. 9
  • Using authentic case reports does not fully resolve realism concerns, while privacy makes fully authentic counseling data nearly impossible to obtain and share ethically.Source: Limitations · p. 9
  • Cost and reproducibility constraints limit the availability of stronger solutions to potential bias in LLM evaluation.Source: Limitations · p. 9
  • Multi-turn memory raises client-privacy concerns, and LLMs may provide harmful advice or incorrect responses.Source: Limitations · p. 9
  • The authors characterize the work as exploratory rather than a completed, mature, deployable study and will not promote it directly as a commercial service to avoid harming vulnerable populations.Source: Limitations · p. 9

Ethical Statement

The section describes safeguards for participants and a staged process for constructing multi-session counseling data, organizing session information, and evaluating conversation quality.

Ethical Statement

  • Case reports were anonymized under institutional review approval; participants gave informed consent, could withdraw, and received compensation, while human evaluation remained limited in scale.The authors note that human expert evaluations are used to mitigate potential LLM-evaluation bias, but cost and reproducibility considerations constrain their scale.Source: Ethical Statement · p. 9

A Client Profile and Counseling Goals Construction

B Seed Conversation Construction

  • Multiple brief seed conversations use the client profiles and session goals, are generated simultaneously for cross-session coherence, and guide the complete conversations.The seeds are intended to help sessions flow logically from earlier sessions and set up later ones.Source: B Seed Conversation Construction · p. 11

C Counseling Generation

  • Seed conversations are expanded into natural-sounding complete sessions designed to maintain contextual coherence and resemble human counseling language and techniques.The process targets both continuity across sessions and natural-sounding counseling.Source: C Counseling Generation · p. 11

D Memory Generation

  • Counseling-session memories summarize key information in a structured format to support progress tracking, conversational consistency, and recall of earlier sessions.The memory is based on the counseling conversation and organizes extracted session information.Source: D Memory Generation · pp. 11–12

E Data Evaluation

  • Data quality is assessed with conversational metrics and the WAI-SF, which measures therapeutic relationship across Goal Agreement, Task Agreement, and Bond.The conversational dimensions are coherence, guidance, helpfulness, and empathy.Source: E Data Evaluation · p. 12

F Experiment Details

The experiments use consistent fine-tuning settings and a simulated client across comparisons, with PANAS scores evaluated after each counseling session. The single-session model performs better than some baselines but does not match multi-session models.

F Experiment Details

  • Meta-Llama-3-8B-Instruct was fine-tuned for two epochs on SMILE, Cactus, and MusPsy-Dataset using LoRA with rank 32, alpha 16, and learning rate 2e-4.Training used the Llama-Factory library.Source: F Experiment Details · p. 13

G LLM Client Construction

  • Comparisons maintain the same simulated client from the initial session, and some Chinese materials are translated into English to align the language context.The passage also begins describing steps to make the simulated interaction more realistic and continuous.Source: G LLM Client Construction · p. 13
  • MusPsy-Model (single) performs better than some baselines but does not produce the same effects as multi-session models.Figure 17 presents emotional changes for the LLM client and performance changes for the LLM counselor across multiple sessions.Source: Figure 17 · p. 14
  • After each session, an LLM evaluates the likely client's PANAS score.The evaluation prompt is shown in Figure 19.Source: G LLM Client Construction · p. 14

H Single Session Experiments

The single-session experiment distinguishes the effects of the dataset from those of the model design: Task 3 alone cannot track client information or dynamically adjust goals, while the dataset and full design make distinct contributions.

H Single Session Experiments

  • Task 3 alone cannot track client information or dynamically adjust counseling goals; its performance significantly decreases, though it still outperforms some baselines.The experiment informed the model only of the current consultation session number.Source: H Single Session Experiments · p. 14
  • The authors attribute the single-session model's remaining baseline advantage to MusPsy-Dataset's richer content, including more advanced counseling techniques and goals.This is the authors' attribution, rather than a separately reported causal test.Source: H Single Session Experiments · pp. 14–15
  • The authors describe their contributions as encompassing both the dataset and the model design.The discussion points to Figure 17 in support of this multifaceted contribution.Source: H Single Session Experiments · p. 15

Related papers

Loading 2506.06626v1…