PsyCounAssist: AI-Powered Counseling System

• PsyCounAssist is an AI-powered counseling assistant that integrates advanced emotion tracking and proactive risk management through dual-mode (Agent and LLM) inference.
• It combines modular multi-agent architectures with end-to-end chain-of-thought pipelines to align counselor responses with client emotion trajectories.
• The system utilizes interactive role-playing data generation and rigorous safety protocols to ensure robust, adaptive, and clinically effective counseling support.

PsyCounAssist is an AI-powered psychological counseling assistant designed to provide empathic, safe, and clinically relevant support through LLMs. The system strategically integrates modules for fine-grained emotion tracking, proactive risk management, and comprehensive evaluation, combining modular multi-agent architectures with end-to-end chain-of-thought (CoT) LLM pipelines. PsyCounAssist operationalizes contemporary counseling theory by tracking client emotion trajectories, aligning the counselor's responses with these dynamics, and enforcing risk mitigation protocols at every conversational turn, setting new benchmarks for emotional insight, safety, and therapeutic effectiveness within automated counseling systems (Xia et al., 18 Jan 2026).

1. Core System Architecture and Inference Paradigms

PsyCounAssist is structured around the dual-module PsychēChat framework: an Emotion Management Module and a Risk Control Module. Each module is instantiated in two system inference paradigms:

• Agent Mode: Explicit multi-agent pipeline, where distinct LLMs (Emotion Tracker, EFT Counselor, Safety Analyzer, Dialogue-Guided Seeker) are called sequentially via function-calling APIs.
• LLM Mode: A single LLM infers all pipeline stages in a unified chain-of-thought, reducing inference latency and simplifying integration.

Workflow Sequence:

1. Emotion Tracking Agent: Given the current client utterance and dialogue history, predicts a primary and up to two secondary emotions utilizing the Plutchik Wheel at three intensity levels; annotates emotion shift $\Delta E_t = E_t - E_{t-1}$.
2. Emotion Memory ($\mathcal{M}$): Maintains per-turn emotion labels, enabling shift-tracking and cumulative trajectory analysis.
3. EFT Draft Response: Counselor module generates a draft response using staged therapeutic components (explore/comfort/action).
4. Dialogue-Guided Seeker Agent: Simulates $k$ potential next-turn client responses.
5. Safety Analysis Agent: Assigns each simulated next-turn a risk score $R(\hat u_{t+1}^{(i)})$; if any $R > \theta$ (default $\theta=0.5$), triggers response regeneration or supervisory intervention (Xia et al., 18 Jan 2026).

A summarized agent-vs-LLM mode comparison is provided below:

Model	Sentient↑	Success↑	Failure↓	Inference
Qwen3-8B + Agent Mode	78.01	51.00%	11.00%	slow
Qwen3-8B + LLM Mode	69.92	42.33%	16.33%	fast

Agent Mode is favored for interpretability and maximum oversight, whereas LLM Mode delivers higher throughput for lower-risk or production deployments.

2. Interactive Role-Playing Data Generation

A critical prerequisite for PsyCounAssist is a data synthesis pipeline using interactive LLM-based role-play:

• Seed corpus: 3k–5k multi-turn counseling transcripts (e.g., PsyDT corpus).
• Role Card Extraction (Gemini-2.5-flash): For each session, extract demographic metadata, Big-Five personality profile, inner monologue, emotion/advice goals, initial emotion, and annotated event triggers.
• Synthetic Dialogue Generation:

1. GPT-4.1-mini acts as the “seeker”, producing turn $(u_t, e_t, \text{trigger})$ in adherence to its profile.
2. Emotion Tracking Agent updates emotion state and memory.
3. EFT Counsel Agent generates a multi-stage draft.
4. Dialogue-Guided Seeker simulates possible next-client responses.
5. Safety Analysis Agent scores all simulated responses for risk, triggering re-generation if thresholds are crossed.

Annotated outputs per turn include: discrete emotion labels, emotion shift/cause, turn-level safety risk score $R_t$, and dialogue outcome filters (final emotion non-negative; at least one goal achieved). The resulting dataset (PsychēDialog) contains over 1,000 full dialogues, average 13.1 turns, used for both agent-mode and LLM-mode training (Xia et al., 18 Jan 2026).

3. Emotion Management and Shift Tracking

PsyCounAssist employs Plutchik’s model, quantizing emotion intensity to $E_t \in [-3, 3]$:

• Emotion shift: $\Delta E_t = E_t - E_{t-1}$
• Emotion Improvement Score (EIS): $\mathcal{M}$0, tracks cumulative emotional relief.
• Emotion Degradation Score (EDS): $\mathcal{M}$1, quantifies average negative dips.

These metrics underpin both automation (e.g., effect monitoring) and structured human evaluation via the PsychēEval protocol, supporting robust comparisons between models on emotional insight.

4. Risk Control and Safety Mechanisms

Risk detection and mitigation are integral at both the turn and session levels:

• Seeker simulation: Given counselor output $\mathcal{M}$2, simulate $\mathcal{M}$3 possible client replies.
• Risk scoring: $\mathcal{M}$4 per simulated reply.
• Turn-level: $\mathcal{M}$5; flag or regenerate if $\mathcal{M}$6.
• Session-level: $\mathcal{M}$7 supports audit and supervisor alerting.
• Safety monitoring: Real-time logging of emotion/risk metrics; integration of crisis-hotline escalation and human-in-the-loop fallback for high-risk states (Xia et al., 18 Jan 2026).

5. Training, Evaluation, and Performance

Model Training:

• Agent Mode: Fine-tune three specialized Qwen-based LLMs on module-specific datasets and one Dialogue-Guided Seeker agent; function-calling APIs orchestrate pipeline.
• LLM Mode: Fine-tune Qwen or comparable LLM on full dialogues with appended chain-of-thought structure.

Recommended hyperparameters include 3 epochs, learning rate $\mathcal{M}$8, 5% warm-up, temperature 0.0.

Evaluation Protocols:

• SAGE: Measures sentient agent’s emotion improvement curve, success/failure rates.
• ESC-Eval: Human/LLM scoring of fluency, empathy, skill, overall quality.
• PsychēEval: Automatic EIS, EDS, Goal Achievement Rate (GAR), RLS risk metrics; human 0–5 scoring on empathy, professionalism, effectiveness, safety.

Key results:

• PsychēChat-Agent: EIS 6.44 (highest), EDS 0.61, GAR 0.94, RLS 0.0062 (lowest), human empathy 4.54/5, professionalism 4.23/5, effectiveness 4.45/5, safety 4.41/5 (best among benchmarks) (Xia et al., 18 Jan 2026).

6. Safety, Deployment, and Practical Considerations

PsyCounAssist includes operational safeguards and deployment recommendations:

• Inference safety checks: Compute $\mathcal{M}$9 each turn; if $k$0, flag for review or demand regeneration.
• API exposure: Provide emotion/risk APIs (emotion_tool(), safety_tool()) for integration with supervisory dashboards (e.g., “thermometer”, “risk meter” UI widgets).
• Human review triggers: For high-risk states or model confidence below threshold, escalate to human specialists.
• Data-adaptive retraining: Periodically re-fine-tune models on new, user-consented deployments to adapt language/cultural patterns.
• Deployment mode: Use LLM Mode for low-latency production; Agent Mode for high-risk or interpretability-centric research scenarios (Xia et al., 18 Jan 2026).

7. Data Requirements and Recommendations

Successful deployment of PsyCounAssist requires:

• Seed library: Minimum 3,000–5,000 multi-turn dialogues, augmented with comprehensive role-card (demographics, Big Five, goals) annotation.
• Synthesis pipeline: Generate $k$11,000 full dialogues with per-turn emotion and risk labels.
• Fine-tuning strategy: Three agent modules plus seeker simulation for Agent Mode; integrated CoT fine-tuned LLM for LLM Mode.
• Continuous monitoring: Log emotion and risk trajectories, support audit and supervision, and provide immediate escalation pathways.

By adopting this protocol, PsyCounAssist achieves robust, interpretable, and adaptive counseling assistance with state-of-the-art metrics in emotional insight and safety risk control, validated under rigorous experimental and real-world conditions (Xia et al., 18 Jan 2026).