Auditing Stealth Sycophancy in Mental-Health Dialogue: Structured Clinical-State Diagnostics and Clean Matched Benchmarks
Not on the map. Why?
Abstract (arXiv)
Mental-health dialogue models are increasingly evaluated by AI-based evaluators, yet these evaluators often treat surface empathy, supportiveness, or fluency as evidence of safety. In this paper, we study a hidden failure mode that we call implicit sycophancy: a response may appear empathetic while implicitly reinforcing catastrophizing, avoidance, hopeless prediction, or CBT-style labeling. To examine this problem, we introduce a diagnostic benchmark for implicit-sycophancy detection, built from three representative mental-health dialogue sources covering everyday peer support, counseling-style emotional support, and crisis-oriented interaction, and further construct a leakage-audited clean single-response matched benchmark with 500 contexts and 1,500 matched response windows. We then propose Dynamic Emotional Signature Graphs (DESG), a structured offline audit framework that separates LLM-based state extraction from final scoring and evaluates clinical direction through semantic, affective, and cognitive-distortion state transitions rather than free-form LLM judgment. Unlike metadata, surface-style, lexical, embedding, and rubric-LLM baselines, DESG scores the direction of clinical-state change induced by a response; on the leakage-audited clean matched benchmark, DESG-StateRisk improves over the strongest non-DESG baseline by 0.0488 macro-F1 and achieves the best harmful-risk detection result. These results suggest that evaluating implicit sycophancy requires explicit clinical-state modeling together with leakage checks, shortcut controls, and competitive baselines.
Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Builds on 28 of its 43 references that are in the map. First posted on arXiv 5 May 2026.
Builds on
28 of 43 references.
References · 28
- 2019Sentence-BERT: Sentence Embeddings using Siamese BERT-Networkscited by 9,049
- 2019BERTScore: Evaluating Text Generation with BERTcited by 5,462
- 2024A Survey on LLM-as-a-Judgecited by 1,384
- 2024LLM Evaluators Recognize and Favor Their Own Generationscited by 651
- 2024From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judgecited by 480
- 2024Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Modelscited by 431
- 2024Humans or LLMs as the Judge? A Study on Judgement Biasescited by 314
- 2021Towards Emotional Support Dialog Systemscited by 294
- 2024Judging the Judges: A Systematic Study of Position Bias in LLM-as-a-Judgecited by 249
- 2024Self-Preference Bias in LLM-as-a-Judgecited by 248
- 2023Generative Judge for Evaluating Alignmentcited by 163
- 2024OffsetBias: Leveraging Debiased Data for Tuning Evaluatorscited by 94
16 more
- 2024CPsyCoun: A Report-based Multi-turn Dialogue Reconstruction and Evaluation Framework for Chinese Psychological Counselingcited by 88
- 2023Split and Merge: Aligning Position Biases in LLM-based Evaluatorscited by 71
- 2024PARIKSHA: A Large-Scale Investigation of Human-LLM Evaluator Agreement on Multilingual and Multi-Cultural Datacited by 37
- 2024PsyDT: Using LLMs to Construct the Digital Twin of Psychological Counselor with Personalized Counseling Style for Psychological Counselingcited by 37
- 2024Are LLM-Judges Robust to Expressions of Uncertainty? Investigating the effect of Epistemic Markers on LLM-based Evaluationcited by 29
- 2025CounselBench: A Large-Scale Expert Evaluation and Adversarial Benchmarking of Large Language Models in Mental Health Question Answeringcited by 20
- 2024Understanding the Therapeutic Relationship between Counselors and Clients in Online Text-based Counseling using LLMscited by 19
- 2024Do Large Language Models Align with Core Mental Health Counseling Competencies?cited by 19
- 2025Crowd Comparative Reasoning: Unlocking Comprehensive Evaluations for LLM-as-a-Judgecited by 18
- 2025TRACT: Regression-Aware Fine-tuning Meets Chain-of-Thought Reasoning for LLM-as-a-Judgecited by 16
- 2024ESC-Eval: Evaluating Emotion Support Conversations in Large Language Modelscited by 15
- 2025YESciEval: Robust LLM-as-a-Judge for Scientific Question Answeringcited by 15
- 2024CBT-Bench: Evaluating Large Language Models on Assisting Cognitive Behavior Therapycited by 12
- 2025A Survey of Large Language Models in Psychotherapy: Current Landscape and Future Directionscited by 12
- 2025Feel the Difference? A Comparative Analysis of Emotional Arcs in Real and LLM-Generated CBT Sessionscited by 11
- 2024MentalGLM Series: Explainable Large Language Models for Mental Health Analysis on Chinese Social Mediacited by 9
Built on by
Citing papers per year: 2026: 1.
1 paper.