Skip to content

PsychēChat: An Empathic Framework Focused on Emotion Shift Tracking and Safety Risk Analysis in Psychological Counseling

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/XNeil12138/PsycheChat
Area: LLM Reasoning
Keywords: Psychological Counseling, Emotion Shift Tracking, Safety Risk Analysis, Emotion-Focused Therapy, Dual Inference Paradigms

TL;DR

PsychēChat integrates Emotion-Focused Therapy (EFT) to explicitly model seekers' cross-turn emotion shift trajectories and anticipate potential seeker reactions for proactive safety risk prevention, synthesizing 1,003 multi-turn counseling dialogues (PsychēDialog) and providing both multi-agent collaborative and end-to-end long chain-of-thought inference paradigms that outperform existing psychological LLMs on SAGE and ESC-Eval benchmarks.

Background & Motivation

In recent years, the global demand for mental healthcare has surged dramatically, with depression, anxiety, and chronic stress emerging as predominant public health burdens. However, the scarcity of licensed psychotherapists, combined with prohibitive costs and scheduling friction in conventional clinical therapy, prevents the vast majority of distressed individuals from obtaining timely, consistent, and affordable professional care. The rapid evolution of large language models (LLMs) has inspired specialized counseling dialogue systems such as SoulChat, MeChat, and CPsyCoun to offer automated emotional support. Nevertheless, existing systems treat emotional comprehension as a single-turn, momentary snapshot, relying heavily on generic sympathetic affirmations and failing to capture dynamic cross-turn emotional evolutions across full counseling sessions.

In classical clinical psychology, Emotion-Focused Therapy (EFT) emphasizes that emotional assessment is an ongoing, moment-by-moment process centered on the trajectory of emotional states that individuals enter, remain stuck in, or transition through. When a distressed seeker shifts from initial self-doubt ("Am I a failure?") to temporary rationalization ("Maybe it's just for now") and ultimately plunges into profound existential devaluation ("Am I even worth believing in?"), responding solely to the isolated momentary utterance with generic comfort leaves counseling superficial, failing to address the core unfulfilled psychological need to be affirmed and valued. Concurrently, psychological counseling carries acute safety sensitivities; cold, dismissive, or purely intellectualized responses frequently escalate resistance or trigger emotional dysregulation. Most existing systems rely on passive dataset filtering and implicit training alignment, lacking proactive mechanisms to anticipate how a seeker might react before finalizing a response, thereby failing to uphold the foundational ethical principle of nonmaleficence.

The fundamental tension lies in the necessity for deep therapeutic intervention, which demands fine-grained cross-turn emotional trajectory tracking to unlock core unfulfilled needs, while penetrating emotional interventions simultaneously heighten the risk of triggering adverse emotional breakdowns if deployed without anticipatory safeguards. To overcome this limitation, PsychēChat introduces explicit cross-turn emotion shift tracking and anticipatory safety risk prevention grounded in EFT principles, synthesized via interactive role-playing to yield the PsychēDialog benchmark, and operationalized through both an interpretable multi-agent pipeline and an efficient end-to-end long chain-of-thought paradigm. Core idea: model psychological counseling as an explicit loop of "emotion shift tracking \(\to\) therapeutic planning \(\to\) anticipatory reaction simulation and safety interception," operationalizing EFT principles to safely transform maladaptive stuck emotions while strictly enforcing clinical safety boundaries.

Method

Overall Architecture

The PsychēChat architecture operates across two integrated dimensions: interactive role-playing dialogue synthesis during data construction and dual inference paradigms during model deployment. During dialogue synthesis, structured seeker role cards (encompassing Big Five traits, life distress, dual goals, and event-triggered emotional dynamics) engage in multi-turn interactions with a three-stage counselor pipeline. Upon receiving a seeker utterance, the counselor pipeline first executes the Emotion Management Module: an Emotion Tracking Agent identifies primary/secondary emotions, analyzes direct triggers and attitudes, and updates a persistent Emotion Memory to track short-term shifts and macro trends. Next, an EFT Counselor Agent leverages the three stages of Helping Skills (exploration, insight, action) and EFT principles to select therapeutic strategies and formulate a draft response. Crucially, the Risk Control Module deploys a Dialogue-Guided Seeker Agent to simulate potential non-compliant seeker reactions (silence, extreme sensitivity, or explosive anger), which a Safety Analysis Agent audits across clinical risk dimensions. Detected risks trigger revision guidelines and iterative regeneration (up to 5 attempts). During deployment, this pipeline is configured into either modular agent-based tool calls (Agent Mode) or distilled into a unified first-person long chain-of-thought (LLM Mode).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    In["Seeker Utterance & Context<br/>Dialogue History Input"] --> EM["Emotion Shift Tracking & Dynamic Memory Updating<br/>Identify Primary/Secondary Emotions + Trace Causes + Update Emotion Memory"]
    EM --> EFT["EFT Counselor Strategy Selection & Draft Generation<br/>Exploration/Insight/Action Stages + Core Emotion Processing Guidelines"]
    EFT --> RC["Anticipatory Safety Risk Interception & Iterative Revision<br/>Simulate Silence/Sentimental/Anger Reactions + Multi-dimension Risk Audit"]
    RC -->|Risk Detected / Max 5 Retries| EFT
    RC -->|Safety Audit Passed| Dual["Modular Multi-Agent Pipeline vs Unified Long CoT Paradigms<br/>Tool-Calling Agent Mode / End-to-End Distilled CoT LLM Mode"]
    Dual --> Out["Final Empathetic & Safe Counselor Response<br/>Context-Aware & Risk-Mitigated Output"]

Key Designs

1. Emotion Shift Tracking & Dynamic Memory Updating: Transitioning from Static Snapshots to Cross-Turn Trajectory Modeling Prior counseling models rely on static, turn-isolated emotion tagging that overlooks emotional progression and underlying psychological etiology. PsychēChat adopts Plutchik’s Wheel of Emotions, categorizing emotional states into eight fundamental emotion groups across three discrete intensity tiers. In each dialogue round, the Emotion Tracking Agent extracts one primary emotion and up to two secondary emotions from the seeker's latest utterance, dissecting immediate environmental triggers and the seeker's subjective stance toward their feelings. Concurrently, an external Emotion Memory retains the historical sequence of emotional states to derive recent shift trajectories and global emotional trends. By identifying exact moments of qualitative emotional transition and analyzing underlying core psychological drivers, the counselor uncovers deep unfulfilled needs that guide substantive emotional transformation.

2. Anticipatory Safety Risk Interception & Iterative Revision: From Passive Filtering to Proactive Defense In clinical mental health support, poorly calibrated counselor interventions can precipitate severe therapeutic rupture or acute distress. Instead of relying solely on post-hoc static data hygiene, PsychēChat institutes an active anticipatory verification loop. The Dialogue-Guided Seeker Agent—deliberately isolated from the seeker's private role card to mirror real-world uncertainty—generates three plausible non-compliant seeker responses conditioned on dialogue history and the counselor's draft: withdrawal/silence, excessive sentimental vulnerability, or explosive anger. The Safety Analysis Agent evaluates these simulated outcomes against comprehensive safety criteria, assessing self-harm ideation, emotional escalation, behavioral danger signals, and relationship deterioration. When risks are detected, the agent formulates concrete emotional containment and risk-reduction directives, requiring the EFT Counselor Agent to iteratively revise its response (allowing up to five regeneration cycles) or discard unsafe dialogue samples entirely.

3. Modular Multi-Agent Pipeline vs Unified Long CoT Paradigms: Reconciling Interpretability with Inference Latency To bridge the gap between rigorous multi-agent validation and real-world deployment efficiency, PsychēChat introduces two complementary modeling paradigms: - Agent Mode: Faithfully mirrors the multi-role synthesis architecture by decoupling Emotion Tracking, EFT Counseling, and Safety Analysis into distinct agents managed via structured <tool_call> and <tool_response> function calling. The agents are trained jointly, providing full modular transparency, explicit intermediate auditing, and high clinical interpretability. - LLM Mode: Compresses the multi-stage reasoning pipeline into a coherent, first-person internal long chain-of-thought (CoT) prefixed to the final response. The reasoning path sequentially executes: (1) Emotion Shift Tracking, (2) Current Counseling Planning, (3) Safety Risk Analysis, and (4) Integration and Response. Fine-tuned with this internalized CoT, a single backbone executes end-to-end inference in a single forward pass, reducing deployment GPU memory requirements by 50% and cutting per-turn inference latency by 62.7% (from 7.44s to 2.78s on an RTX 4090).

Loss & Training

The authors fine-tuned Qwen2.5-7B-Instruct and Qwen3-8B across both Agent Mode and LLM Mode datasets using the lightweight training framework ms-swift. The models were optimized for 3 epochs using standard supervised fine-tuning (SFT) with a learning rate of \(1 \times 10^{-5}\) and a warmup ratio of 0.05. During inference and benchmarking, the decoding temperature was set strictly to 0.0 to eliminate stochastic variance and ensure deterministic reproducibility.

Key Experimental Results

Main Results

The models were evaluated across three comprehensive frameworks: SAGE (an interactive multi-turn benchmark evaluating emotional elevation and social cognition with a Sentient Agent up to 10 turns, where scores \(>100\) denote success and \(<10\) denote failure), ESC-Eval (multi-dimensional assessment of empathy, skill, fluency, and overall dialogue quality), and the custom PsychēEval benchmark.

Model Category Model Name SAGE Sentient ↑ SAGE Success(%) ↑ SAGE Failure(%) ↓ ESC-Eval Emp. (0-10) ↑ ESC-Eval Avg. (0-10) ↑
Closed-Source LLMs GPT-4o \(54.06 \pm 0.69\) \(20.67 \pm 5.56\) \(25.00 \pm 3.74\) 10.00 9.66
Closed-Source LLMs DeepSeek-V3 \(76.99 \pm 2.69\) \(47.67 \pm 6.85\) \(9.00 \pm 0.82\) 10.00 9.83
Closed-Source LLMs Qwen-MAX \(62.88 \pm 3.75\) \(31.00 \pm 1.41\) \(16.33 \pm 2.87\) 10.00 9.82
Psychological LLMs SoulChat \(5.22 \pm 0.19\) \(0.00 \pm 0.00\) \(93.33 \pm 1.25\) 8.00 8.01
Psychological LLMs MeChat \(14.62 \pm 1.89\) \(1.67 \pm 0.94\) \(76.00 \pm 0.82\) 9.08 8.86
Psychological LLMs CPsyCounX \(10.08 \pm 2.03\) \(0.67 \pm 0.47\) \(80.00 \pm 3.74\) 8.89 8.60
Psychological LLMs SoulChat2.0 \(60.06 \pm 2.11\) \(24.33 \pm 2.87\) \(17.33 \pm 0.47\) 9.92 9.47
Psychological LLMs + PsyAdvisor \(57.67 \pm 1.36\) \(16.67 \pm 1.70\) \(17.00 \pm 1.63\) 8.59 8.07
Psychological LLMs SoulChat-R1 \(70.87 \pm 5.08\) \(42.67 \pm 18.2\) \(16.33 \pm 1.25\) 9.74 9.04
Ours Qwen2.5-7B + PsychēChat-LLM \(69.55 \pm 1.53\) \(42.33 \pm 4.11\) \(15.33 \pm 3.30\) 9.99 9.64
Ours Qwen2.5-7B + PsychēChat-Agent \(71.80 \pm 1.05\) \(44.33 \pm 6.24\) \(15.33 \pm 3.30\) 9.99 9.57
Ours Qwen3-8B + PsychēChat-LLM \(69.92 \pm 2.25\) \(42.33 \pm 2.62\) \(16.33 \pm 3.86\) 9.85 9.15
Ours Qwen3-8B + PsychēChat-Agent 78.01 ± 1.60 51.00 ± 9.27 11.00 ± 2.83 10.00 9.51

Under PsychēEval (evaluated via automated metrics EIS, EDS, GAR, RLS and double-blind human ratings by three licensed clinical psychologists on a 0–5 scale):

Method Auto EIS ↑ Auto EDS ↓ Auto GAR ↑ Auto RLS ↓ Human Empathy ↑ Human Professionalism ↑ Human Effectiveness ↑ Human Safety ↑
SoulChat2.0 2.65 0.30 0.93 0.0185 \(3.85 \pm 0.16\) \(3.53 \pm 0.12\) \(3.87 \pm 0.18\) \(3.82 \pm 0.08\)
SoulChat-R1 2.59 0.29 0.90 0.0085 \(3.85 \pm 0.28\) \(3.49 \pm 0.28\) \(3.81 \pm 0.33\) \(3.86 \pm 0.04\)
PsychēChat-LLM 2.53 0.26 0.86 0.0061 \(3.85 \pm 0.35\) \(3.33 \pm 0.37\) \(3.87 \pm 0.31\) \(4.05 \pm 0.35\)
PsychēChat-Agent 2.67 0.27 0.94 0.0062 4.54 ± 0.34 4.23 ± 0.52 4.45 ± 0.32 4.41 ± 0.30

Ablation Study

The ablation investigation examines the individual contributions of Emotion Management (EM) and Risk Control (RC) during dialogue synthesis, training Qwen3-8B under Agent Mode. A cross-scenario consistency metric (Stability, standard deviation of Sentient scores) was included, where lower values indicate greater cross-context consistency:

Configuration SAGE Sentient ↑ SAGE Stability ↓ SAGE Success(%) ↑ SAGE Failure(%) ↓ ESC-Eval Flu. ESC-Eval Emp. ESC-Eval Ovl. ESC-Eval Avg.
PsychēChat-Agent (Full) 78.01 ± 1.60 32.04 ± 1.21 \(51.00 \pm 9.27\) \(11.00 \pm 2.83\) 10.00 10.00 9.99 9.51
w/o EM (no emotion tracking) \(71.90 \pm 1.28\) \(34.13 \pm 0.65\) \(39.00 \pm 2.16\) \(12.00 \pm 2.16\) 9.98 9.99 9.69 9.29
w/o RC (no risk control) \(76.15 \pm 1.50\) \(33.93 \pm 0.71\) 52.00 ± 7.48 \(11.67 \pm 1.25\) 9.96 9.98 9.95 9.59
w/o EM & RC (neither module) \(71.68 \pm 0.82\) \(34.27 \pm 0.40\) \(41.67 \pm 6.94\) \(11.67 \pm 1.25\) 9.98 9.99 9.90 9.55
SoulChat2.0-1003 (matched seed data) \(57.29 \pm 0.24\) \(36.72 \pm 0.48\) \(23.00 \pm 2.16\) \(20.33 \pm 0.47\) 9.57 9.49 9.38 9.13

Key Findings

  1. Emotion Management drives therapeutic elevation: Removing EM precipitates a sharp decline in SAGE Sentient (\(78.01 \to 71.90\)) and a 12.0% drop in counseling success rate (\(51.00\% \to 39.00\%\)), proving that explicit tracking of emotional transitions and causal drivers is the primary catalyst for psychological improvement.
  2. Risk Control stabilizes trajectories and suppresses variance: Eliminating RC degrades cross-scenario stability (\(32.04 \to 33.93\)) and increases failure rates. RC prevents overly aggressive emotional probing from inducing secondary distress, ensuring steady, safe emotional de-escalation.
  3. Role-playing synthesis yields superior data quality: On an identical scale (1,003 dialogues) and topic distribution, the unaugmented baseline (w/o EM & RC, scoring 71.68 on Sentient) dramatically outperforms SoulChat2.0-1003 (57.29), highlighting the intrinsic benefit of multi-turn role-playing over static seed dialogues.
  4. Agent vs. LLM Mode trade-off: Agent Mode achieves peak counseling efficacy and human clinical approval (Sentient 78.01, Human Professionalism 4.23), while LLM Mode delivers a 62.7% reduction in latency (2.78s per turn) on a single 24GB GPU, offering an ideal trade-off for real-time applications.

Highlights & Insights

  • Explicit Emotion Trajectory Tracking Grounded in EFT: Grounding counseling interactions in Plutchik's emotion space and Emotion-Focused Therapy transforms the dialogue system from an echo-chamber of shallow sympathy into an attuned therapeutic partner capable of resolving underlying maladaptive emotional schemas.
  • Anticipatory Reaction Simulation as Active Risk Interception: Instead of relying exclusively on reactive toxicity checks, the framework actively simulates potential extreme seeker responses (silence, extreme sensitivity, explosive anger) prior to output generation, intercepting therapeutic blunders before they reach the user.
  • Distillation of Multi-Agent Interactions into Unified Long CoT: The compilation of multi-agent deliberative reasoning into a single first-person chain-of-thought bridges the long-standing divide between high-fidelity agentic reasoning and low-latency production serving.

Limitations & Future Work

  • High Computational Overhead in Synthesis: Interactive multi-agent synthesis requires roughly 100 API calls (~240K tokens) per dialogue, restricting PsychēDialog to 1,003 dialogues across 12 counseling domains and posing scalability constraints.
  • Simulated vs. In-the-Wild Clinical Discrepancy: While validated by licensed clinical psychologists on 50 sampled cases, the primary experimental benchmarks remain simulated; real-world longitudinal effectiveness in clinical psychiatric environments requires deeper validation.
  • Cultural Specificity: The personas, narrative contexts, and linguistic styles in PsychēDialog are presently grounded in Chinese cultural scenarios; cross-cultural transferability of role cards and therapeutic phrasing warrants further research.
  • vs SoulChat / MeChat: Early counseling models expanded single-turn Q&A datasets into conversational threads without explicit tracking of cross-turn emotional arcs; PsychēChat incorporates structured emotion trajectories and causal attribution to achieve significantly deeper therapeutic impact.
  • vs CPsyCoun / CATCH: While CPsyCoun and CATCH introduce consultation reports or Single-Session Therapy stages, they employ static multi-stage generation; PsychēChat relies on dynamic role-playing with explicit risk simulation.
  • vs PsyAdvisor / PsyGUARD: PsyGUARD operates offline suicide risk detection and PsyAdvisor optimizes question planning; PsychēChat unifies emotion progression tracking, strategy selection, reaction simulation, and iterative safety remediation into a closed-loop defense architecture.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ First framework to systematically integrate cross-turn emotion shift tracking with anticipatory reaction-driven risk control in psychological LLMs.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive validation spanning SAGE, ESC-Eval, custom PsychēEval metrics, detailed ablations, and double-blind clinical evaluations by licensed psychologists.
  • Writing Quality: ⭐⭐⭐⭐⭐ Exemplary synthesis of clinical psychology principles (EFT) with modern LLM system design, supported by exhaustive empirical metrics and clear diagrams.
  • Value: ⭐⭐⭐⭐⭐ Establishes a foundational, safety-guaranteed paradigm for developing responsible, high-empathy mental healthcare conversational agents.