Beyond the Response: Examining Reasoning and Execution Fidelity in Large Language Models for Mental Health
Year: 2026
Type: conference-paper
Abstract: Large language models (LLMs) are increasingly used for mental health support, raising questions about how they reason, communicate, and support accountability in safety-critical contexts. Existing evaluations primarily assess final outputs, such as accuracy or perceived empathy, without examining how responses are internally planned or whether stated plans are upheld in execution. We conduct a process-level observational audit of execution fidelity, defined as the consistency between commitments stated in self-reported Chain-of-Thought (CoT) and the model's final response. We evaluate five state-of-the-art LLMs on 75 high-risk mental health prompts spanning depression, anxiety, trauma, and emotional distress. For each prompt, we analyze both the model's self-generated Chain-of-Thought (CoT) and its final response, treating CoTs as inspectable planning artifacts. Using a mixed-methods approach, we qualitatively code the response commitments articulated during planning, such as empathy, guidance, boundary setting, and problem framing, and quantitatively assess whether these commitments are upheld in execution using lexical analysis and natural language inference-based alignment checks. Across nearly 400 planning steps derived from these CoTs, we observe systematic differences in how models prioritize commitments during planning, as well as recurrent cases where safety-relevant commitments are omitted in the responses. These include dropped disclaimers, compressed guidance, and substituting generic emotional reassurance. We further show that linguistic instability during planning, such as increased hedging and lexical variability, co-occurs with such commitment violations. Together, our findings position process-level audits as a means of identifying procedural risk and supporting post-deployment oversight of LLMs in mental health contexts.
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166556191