When Seekers Are Hard to Help: Evaluating Emotional Support Dialogue Systems in Worst-Case Interactions¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: Emotional Support Dialogue Systems (ESDS), Worst-Case Evaluation, Seeker Simulator, Emotional Volatility, Robustness Stress Testing
TL;DR¶
Addressing the superficial evaluation ceiling and optimistic bias resulting from cooperative, average-case simulated users in emotional support dialogue systems (ESDSes), this work grounds worst-case interactions through an expert simulation study with licensed counsellors, constructs a controllable worst-case seeker simulator equipped with emotional deterioration, introduces four robustness-oriented evaluation metrics, benchmarks 17 representative models to reveal severe performance drops, and demonstrates that fine-tuning with worst-case synthetic data substantially enhances smaller open-weight models.
Background & Motivation¶
Emotional Support Dialogue Systems (ESDSes) aim to provide empathetic, contextually nuanced, and clinically constructive responses to individuals experiencing acute emotional distress. Driven by the rapid evolution of large language models (LLMs), these systems have demonstrated unprecedented linguistic fluency, conversational flexibility, and accessibility. Given the prohibitive costs, privacy sensitivities, and ethical dilemmas associated with gathering real-world clinical consultation transcripts, contemporary research has predominantly turned toward LLM-simulated seekers for automated benchmark evaluation and pipeline training. However, unconstrained conversational LLM agents inherently manifest normative social biases and sycophantic cooperativeness, consistently defaulting to well-behaved, communicative, and compliant "average-case" personas.
This systemic reliance on cooperative personas induces an acute optimistic bias in dialogue evaluation. In conventional benchmarking scenarios, simulated seekers promptly disclose their core vulnerabilities within the first few turns, respond constructively to system inquiries, readily accept cognitive reframing, and smoothly de-escalate into stable emotional states. Yet, in authentic psychological support contexts, distressed seekers rarely behave as ideal collaborative partners; rather, they frequently display minimal conversational engagement, explicit skepticism, active resistance toward well-intentioned advice, defensive concealment of root issues, erratic emotional volatility, and deeply entrenched cognitive biases. Evaluating support models strictly against cooperative users obscures critical behavioral vulnerabilities and misidentifies polite formulaic agreeableness as authentic clinical efficacy.
Such challenging and resistant encounters are far from rare statistical anomalies; they represent the precise frontier where emotional support systems are most prone to conversational rupture, and where users most critically require genuine clinical patience, nuanced strategy adaptation, and deep affective attunement. A system that appears extraordinarily empathetic in placid interactions frequently degenerates into repetitive generic validation, premature and unsolicited advising, relentless interrogation, or unprincipled sycophancy when confronted with an uncooperative seeker. To bridge this critical methodology gap, this work argues that emotional support systems must be rigorously evaluated not merely against compliant average-case seekers, but under worst-case interactions where seekers are fundamentally hard to help. Core idea: ground worst-case interactions in an expert simulation study with licensed counsellors to identify core unhelpful patterns, build a controllable worst-case seeker simulator incorporating a dynamic emotional deterioration controller and four prompt-level behavioral axes, benchmark 17 mainstream LLMs across generic and four newly formulated robustness metrics, and leverage worst-case synthetic rollouts to dramatically improve the robustness of open-weight supporter models.
Method¶
Overall Architecture¶
The worst-case evaluation framework operationalizes conversational resistance into an empirical, closed-loop simulation and dual-tier assessment system. The overall architecture consists of three interlocking operational stages: 1. Expert Simulation Study & Clinical Grounding: Eight accredited counselling professionals (possessing doctoral training and over 3 years of clinical practice) engage in blind, multi-turn interactions (up to 100 turns) with eight Chinese dialogue systems (seven specialized ESDSes and one general-purpose LLM), followed by 45-minute semi-structured qualitative interviews to construct an empirical taxonomy of difficult seeker behaviors and model failure modes; 2. Worst-Case Seeker Simulator: A modular three-agent pipeline—comprising a Profile Processing Agent, an Emotion Controller, and a Seeker Agent—that dynamically regulates seeker responses via explicit emotional deterioration probability and four orthogonal behavioral constraints; 3. Dual-Tier Metric Evaluation & Training Loop: Incorporating ten standard emotional support metrics alongside four newly formulated worst-case metrics (Deep Emotional Understanding, Guided Exploration, Balanced Emotional Support, and Authentic and Grounded Support) evaluated via zero-temperature Qwen3-Max, culminating in a data synthesis loop where GPT-5.4 generates adversarial interaction traces for LoRA fine-tuning of lightweight open-weight models.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Raw Seeker Profiles & Consulting Situations<br/>Demographic Attributes & Clinical Prompts"] --> B["Profile Processing Agent<br/>GPT-5.4 Normalization of Unstructured Contexts"]
B --> C["Emotion Controller<br/>Predicts Transition Reason, Emotion State & Triggers 0.3 Deterioration"]
C --> D["Seeker Agent Simulation<br/>Generates Response Conditioned on Engagement, Resistance, Style & Disclosure"]
D --> E["Multi-turn Interaction with Evaluated ESDSes<br/>17 Models Tested Across Closed-loop Conversations (Max 20 Turns)"]
E --> F["Dual-Tier Evaluation Protocol<br/>10 Generic Metrics + 4 Worst-case Metrics (DEU, GE, BES, AGS)"]
F --> G["Synthetic Data Fine-Tuning Pipeline<br/>GPT-5.4 Generates Worst-case & Mixed Traces for Small Supporter Models"]
Key Designs¶
1. Empirical Grounding: Six Difficult Seeker Behaviors and Four Supporter Failure Modes
To anchor the conceptual definition of worst-case interactions in real-world clinical psychology rather than subjective prompt hacking, the study conducted in-person simulation sessions with eight licensed counselling professionals. Each expert role-played challenging clients across varied counseling situations and rated system responses using standard scales before reflecting on conversational breakdowns in semi-structured interviews. This qualitative inquiry synthesized six canonical difficult seeker behaviors: - High-Risk Signals: Veiled, indirect, or ambiguous hints of crisis and self-harm; - Low Dialogue Engagement: Terse, perfunctory, or passive one-word contributions; - Resistance to Support: Active rejection, skepticism, and verbal opposition to reassuring attempts; - Limited Self-Disclosure: Strategic withholding of salient background facts, feelings, and core problem contexts; - Emotional Dysregulation: Unstable, highly fluctuating, or escalating negative affective states across turns; - Cognitive Extremity: Rigid, absolutist, catastrophic, or immutable negative cognitive appraisals.
Concurrently, the clinical interviews pinpointed four recurring architectural failures in contemporary conversational systems: failure to detect implicit crises, shallow literalism that misses metaphorical subtext and relational grievances, premature advising prior to problem exploration, and uncritical sycophancy (blindly validating biased viewpoints and reinforcing victim narratives).
2. Worst-Case Seeker Simulator: Tripartite Architecture with Dynamic Affective Deterioration
The automated simulator orchestrates three specialized agents. First, the Profile Processing Agent leverages GPT-5.4 to clean and standardize raw seeker demographic attributes and initial counseling situations extracted from SimPsyDial, eliminating stylistic noise while retaining original psychological intent. Second, the Emotion Controller operationalizes emotional dysregulation and cognitive extremity. At each conversational turn, it conditions on the normalized profile, dialogue history, and prior emotion state to predict an updated emotional state from a standardized vocabulary, producing an explicit Transition Reason and a qualitative Emotion Description. To simulate genuine affective volatility, the controller incorporates a deterioration probability of 0.3: when triggered, it forces the simulated seeker into an intensely negative, defensive, or frustrated state, actively counteracting artificial cooperativeness.
Finally, the Seeker Agent synthesizes the actual conversational response conditioned on the processed profile, emotional status, and four controllable behavioral parameters: - Engagement: Controls conversational verbosity and spontaneous elaboration versus extreme passivity; - Resistance: Dictates the propensity to challenge suggestions, question the supporter's competence, or reject empathy; - Expression Style: Modulates output rhetoric between direct, ambiguous, colloquial, and dismissive vernacular; - Self-Disclosure: Restricts or gates the release of private vulnerability, forcing the supporter to conduct active inquiry.
3. Four Worst-Case-Oriented Evaluation Metrics
While general benchmarks (Ye et al., 2026) evaluate ten standard dimensions (Human-Likeness, Engagement, Empathetic, Personalization, Adaptive Strategies, Consistency, Redundancy, Helpfulness, Mood Improvement, Safety), they fail to detect subtle therapeutic missteps under duress. This paper introduces four dedicated 5-point Likert metrics: - Deep Emotional Understanding (DEU): Measures the system's ability to transcend literal surface syntax and detect underlying feelings of grievance, helplessness, relational friction, or unstated needs masked by resistance; - Guided Exploration (GE): Evaluates whether the supporter effectively deploys open-ended questions, active listening, reflection, and contextual focusing to help under-disclosing seekers elaborate their dilemmas, penalizing premature problem-solving; - Balanced Emotional Support (BES): Assesses the model's capacity to validate raw emotions while maintaining therapeutic neutrality and objective perspective, explicitly penalizing sycophantic reinforcement of destructive, one-sided, or paranoid interpretations; - Authentic and Grounded Support (AGS): Evaluates whether supportive interventions are uniquely grounded in the seeker's idiosyncratic narrative details, penalizing formulaic clichés and empty platitudes (e.g., "I completely understand your pain").
4. End-to-End Evaluation Protocol and Adversarial Synthetic Fine-Tuning
The empirical protocol benchmarks paired interactions generated under identical seeker profiles: an unconstrained Average-Case baseline versus the multi-parameter Worst-Case simulator (capped at 20 turns, allowing early <END> exit if the seeker experiences adequate relief or concludes the system is futile). Scoring is executed via zero-temperature Qwen3-Max, demonstrating robust Spearman correlation with licensed expert ratings (Human-likeness \(\rho=0.71\), Redundancy \(\rho=0.76\), Personalization \(\rho=0.70\)).
In the training pipeline, the highest-performing general model (GPT-5.4) is employed as an expert demonstrator to generate 1,000 average-case and 1,000 worst-case complete multi-turn dialogues over SimPsyDial seeds. Lightweight foundation models (Qwen3-4B-Instruct) are fine-tuned via LoRA (rank = 16, lr = \(2\times 10^{-4}\), 3 epochs) across varying data mixtures (Avg-only, Worst-only, 25/75, 50/50, 75/25, and scaled 2K datasets) to verify whether adversarial synthetic data cures interactional fragility.
Key Experimental Results¶
Main Results¶
The evaluation benchmarks 17 distinct conversational systems across general-purpose commercial LLMs, open-weight lightweight backbones, and specialized Chinese mental health models. Table 1 summarizes performance across key metrics under worst-case interactions, including relative percentage degradation compared to average-case interactions.
| Model Category | Model Name | Human-Likeness (HL) | Engagement (Eng.) | Empathetic (Emp.) | Adaptive Strategies (AS) | Mood Improvement (MI) | Safety (Safe.) | Deep Emotional Understanding (DEU) | Guided Exploration (GE) | Balanced Emotional Support (BES) | Authentic Grounded Support (AGS) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| General LLMs | GPT-5.4 | 4.18 (-16.4%) | 1.82 (-63.2%) | 3.12 (-37.6%) | 3.02 (-39.6%) | 1.52 (-68.9%) | 4.06 (-18.5%) | 2.92 | 2.04 | 3.46 | 2.92 |
| Claude-4.6-Sonnet | 3.96 (-20.8%) | 1.70 (-66.0%) | 2.76 (-44.8%) | 2.40 (-52.0%) | 1.38 (-72.3%) | 4.08 (-18.4%) | 2.20 | 1.92 | 3.10 | 2.28 | |
| DS-V4-Pro | 4.06 (-17.5%) | 1.68 (-65.6%) | 2.88 (-41.5%) | 2.52 (-48.8%) | 1.54 (-67.6%) | 4.24 (-13.8%) | 2.20 | 1.86 | 3.10 | 2.28 | |
| Qwen-3.6-Plus | 4.02 (-19.6%) | 1.68 (-66.0%) | 2.82 (-43.6%) | 2.62 (-47.6%) | 1.50 (-69.6%) | 4.32 (-13.6%) | 2.14 | 1.86 | 3.00 | 2.12 | |
| Doubao-Seed-2.0-Pro | 3.46 (-30.0%) | 1.42 (-67.7%) | 2.32 (-51.7%) | 1.84 (-56.0%) | 1.14 (-71.8%) | 3.82 (-22.4%) | 1.82 | 1.58 | 2.84 | 1.72 | |
| DeepSeek-V3.2 | 3.20 (-36.0%) | 1.36 (-72.2%) | 2.28 (-54.4%) | 1.84 (-62.8%) | 1.10 (-76.9%) | 3.76 (-24.8%) | 2.02 | 1.88 | 3.00 | 1.90 | |
| GPT-4o | 2.62 (-41.3%) | 1.20 (-66.7%) | 2.06 (-51.9%) | 1.46 (-58.5%) | 1.08 (-68.4%) | 3.50 (-28.6%) | 1.76 | 1.50 | 2.88 | 1.44 | |
| Open-Weight | Qwen3-8B | 2.42 (-35.3%) | 1.24 (-56.3%) | 2.08 (-44.1%) | 1.22 (-52.3%) | 1.12 (-55.6%) | 3.54 (-21.0%) | 1.76 | 1.54 | 2.88 | 1.38 |
| Qwen3-4B | 2.18 (-31.4%) | 1.14 (-48.2%) | 2.04 (-36.6%) | 1.10 (-42.7%) | 1.06 (-49.0%) | 3.58 (-12.7%) | 1.74 | 1.42 | 2.86 | 1.32 | |
| Specialized ESDS | SoulChat 2.0 | 2.34 (-48.7%) | 1.12 (-69.4%) | 2.04 (-54.1%) | 1.22 (-67.7%) | 1.02 (-68.5%) | 3.26 (-32.9%) | 1.82 | 1.62 | 2.82 | 1.58 |
| SimPsybot_q | 2.26 (-50.2%) | 1.14 (-68.3%) | 1.88 (-57.1%) | 1.24 (-67.2%) | 1.00 (-68.8%) | 3.02 (-36.8%) | 1.50 | 1.72 | 2.76 | 1.60 | |
| EmoLLM | 2.40 (-40.0%) | 1.12 (-62.2%) | 1.94 (-48.1%) | 1.30 (-54.2%) | 1.00 (-63.0%) | 3.38 (-26.8%) | 1.60 | 1.16 | 2.66 | 1.38 | |
| MindChat | 2.04 (-11.3%) | 1.10 (-32.1%) | 1.46 (-36.0%) | 1.14 (-33.7%) | 1.00 (-23.1%) | 2.68 (-14.6%) | 1.16 | 1.18 | 2.32 | 1.12 | |
| MeChat | 1.90 (-22.1%) | 1.06 (-36.9%) | 1.60 (-39.4%) | 1.00 (-37.5%) | 1.02 (-29.2%) | 2.90 (-16.2%) | 1.20 | 1.14 | 2.46 | 1.08 | |
| CPsyCounX | 1.64 (-13.7%) | 1.06 (-20.9%) | 1.60 (-25.2%) | 1.00 (-16.7%) | 1.00 (-13.8%) | 2.66 (-15.3%) | 1.24 | 1.28 | 2.52 | 1.12 | |
| PsyChat | 2.00 (+3.1%) | 1.12 (-11.1%) | 1.62 (-17.3%) | 1.14 (-1.7%) | 1.06 (-3.6%) | 2.98 (-4.5%) | 1.18 | 1.12 | 2.36 | 1.08 |
Ablation & Fine-Tuning Results¶
To isolate the behavioral factors most lethal to support systems, Table 4 analyzes component ablations on SoulChat 2.0. Concurrently, Table 2 documents the robustness improvements gained by fine-tuning Qwen3-4B-Instruct across various synthetic data curricula.
SoulChat 2.0 Behavioral Component Ablation: - Full Worst-Case Interaction: HL: 2.34 (-48.7%), Eng.: 1.12 (-69.4%), MI: 1.02 (-68.5%), AS: 1.22 (-67.7%); - Emotional Volatility Only: Inflicts the sharpest performance collapse, driving Mood Improvement to 1.10 (-66.0%), Engagement to 1.32 (-63.9%), and Adaptive Strategies down by -63.5%; - Engagement Only: Triggers severe conversational withdrawal (Eng. 2.04, -44.3%; MI 1.92, -40.7%); - Resistance Only: Induces significant refusal penalties (Eng. 2.38, -35.0%; MI 1.90, -41.4%); - Expression Style Only: Reduces fluency and adaptation (HL 3.42, -25.0%; AS 2.44, -35.4%); - Self-Disclosure Only: Exhibits minimal adverse impact (relative declines limited to 7%–11%, with Engagement dropping by only -8.7%).
Qwen3-4B-Instruct Synthetic Data Fine-Tuning Performance (Evaluated under Worst-Case Conditions):
| Fine-Tuning Regimen | Human-Likeness (HL) | Engagement (Eng.) | Empathetic (Emp.) | Adaptive Strategies (AS) | Mood Improvement (MI) | Safety (Safe.) | DEU | GE | BES | AGS |
|---|---|---|---|---|---|---|---|---|---|---|
| Zero-Shot Base Qwen3-4B | 2.72 | 1.18 | 2.10 | 1.28 | 1.10 | 3.50 | 1.68 | 1.36 | 2.80 | 1.44 |
| AvgOnlyFT-1K | 3.80 | 1.50 | 2.60 | 2.14 | 1.30 | 3.50 | 2.28 | 1.78 | 3.14 | 2.22 |
| WorstOnlyFT-1K | 3.88 | 1.52 | 2.76 | 2.50 | 1.26 | 4.00 | 2.46 | 1.86 | 3.26 | 2.46 |
| MixFT-25W/75A-1K | 3.84 | 1.58 | 2.74 | 2.58 | 1.36 | 3.84 | 2.40 | 1.76 | 3.14 | 2.40 |
| MixFT-50W/50A-1K | 3.88 | 1.62 | 2.80 | 2.56 | 1.26 | 3.84 | 2.22 | 1.88 | 3.12 | 2.30 |
| MixFT-75W/25A-1K | 3.94 | 1.66 | 2.84 | 2.50 | 1.38 | 3.98 | 2.28 | 1.86 | 3.22 | 2.22 |
| MixFT-50W/50A-2K | 4.00 | 1.68 | 2.80 | 2.62 | 1.42 | 4.00 | 2.50 | 1.94 | 3.22 | 2.54 |
Key Findings¶
- The Cooperative Illusion and Catastrophic Fragility: While frontier LLMs (GPT-5.4, Claude-4.6, Qwen-3.6-Plus) achieve near-perfect ceiling scores (5.00) under standard cooperative average-case benchmarks (Table 7), their performance plunges precipitously by 60% to 77% in Engagement and Mood Improvement under worst-case interactions. Even the most resilient system fails to sustain constructive engagement or produce meaningful affective recovery when users refuse easy comfort.
- Superiority of General Reasoners over Narrow Specialists: Large general-purpose foundation models exhibit vastly superior resilience compared to specialized clinical dialogue systems. Specialized models like EmoLLM and SoulChat 2.0 suffer absolute degradation to floor levels (approaching 1.0 in Mood Improvement and Engagement), revealing that narrow supervised fine-tuning overfits models to cooperative scripts, causing them to compulsively iterate trite advice ("communicate openly, seek rest") when rejected.
- Rupture Repair and Boundary Reframing Distinguish Tier-1 Models: Qualitative analysis of family conflict interactions (Figure 4) underscores why GPT-5.4 outperforms GPT-4o. When the seeker aggressively dismisses exploratory questioning as useless, GPT-4o loops back to abstract communication ideals, driving the seeker into withdrawal; GPT-5.4 actively acknowledges the procedural mismatch, apologizes, delivers targeted concrete scripts, and when the user despairs that "talking is futile," smoothly pivots from persuading the relative to establishing personal psychological boundaries.
- Complementarity of Adversarial and Cooperative Training Data: Exposing lightweight models to balanced synthetic mixtures (50% worst-case / 50% average-case at 2K volume) elevates baseline performance dramatically across Deep Emotional Understanding (1.68 \(\rightarrow\) 2.50) and Authentic Grounded Support (1.44 \(\rightarrow\) 2.54), proving that worst-case simulation acts as an effective, scalable synthetic generator for resilient alignment.
Highlights & Insights¶
- De-biasing Social Simulation Benchmarks: Formally identifies and challenges the cooperative bias inherent in role-playing LLM evaluations, proving that standard average-case benchmarking rewards superficial eloquence while obscuring severe fragility under conversational friction.
- Clinical Psychology Grounding: Anchors evaluation criteria in rigorous empirical feedback from doctoral-level psychological counsellors, introducing novel target dimensions—such as Balanced Emotional Support (neutrality against sycophantic reinforcement)—that reflect genuine therapeutic ethics.
- Emotional Volatility as Primary Failure Driver: Component ablation decisively demonstrates that unprompted affective escalation and cognitive extremity disrupt dialogue state tracking far more destructively than simple information withholding, delineating the frontier for future dialogic affective computing.
Limitations & Future Work¶
- Simulated Agents Lack Embodied Stakes: Despite rigorous expert calibration and multi-axis prompting, LLM agents remain digital approximations lacking physical embodiment, long-term personal histories, and real-life consequences, limiting direct clinical equivalence.
- Omission of Explicit Suicidal Crisis Scenarios: Due to strict commercial LLM safety filters, the simulator deliberately bypasses generating active suicidal threats or emergency self-harm crises, focusing purely on interpersonal resistance rather than acute psychiatric intervention.
- Cross-Cultural Counseling Generalizability: The empirical baseline relies primarily on Chinese counseling practices and linguistic norms. Validating the universality of resistance behaviors and evaluating Western high-context vs. low-context counseling styles remains an open imperative.
Related Work & Insights¶
- vs. Static Strategy-Classification Benchmarks (ESConv, SMILE): Prior benchmarks evaluate one-step turn-level strategy labeling or generation under cooperative human prompts; this work constructs a dynamic, closed-loop multi-turn testing environment with active behavioral friction.
- vs. LLM Sycophancy & Alignment Research (Ibrahim et al., 2026): While previous work demonstrates that alignment causes models to agree with factual errors or political biases, this paper reveals the therapeutic danger of sycophancy, where models reinforce distorted victim beliefs and validate harmful externalization.
- vs. Standard Simulated Patient Environments (SimPsyDial, EmoHarbor): Unlike existing simulators designed to mirror average seeker personas, this framework introduces an explicit 0.3 affective deterioration probability and adversarial resistance, transforming role-playing from a passive sandbox into an active stress test.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Formulates the first rigorous worst-case evaluation framework for emotional support dialogue systems, integrating expert clinical modeling with dynamic affective deterioration.
- Experimental Soundness: ⭐⭐⭐⭐⭐ Encompasses 8 accredited counseling experts, 17 state-of-the-art models, detailed behavioral ablations, human-judge correlations, and synthetic fine-tuning validation.
- Writing Quality: ⭐⭐⭐⭐⭐ Articulately composed, tightly structured, and seamlessly integrates clinical psychology principles with empirical computational linguistics.
- Impact: ⭐⭐⭐⭐⭐ Exposes the fragile reality behind superficial benchmark success, providing an essential methodology and data generation blueprint for building robust, rupture-resilient conversational agents.