When Less Is More: An Empirical Study of Minimal Responses in Counseling Dialogues and the Behavior of LLMs

In professional psychological counseling, effective support is not always about providing information or advice. Minimal responses—brief utterances like "Mm-hmm," "I see," or "That sounds difficult"—are essential micro-skills. They serve as "continuers" that signal attentive listening and empathy without interrupting the client’s narrative flow. Despite their importance in building a therapeutic alliance, these brief utterances are frequently overlooked in modern dialogue systems and evaluation frameworks, which often prioritize content-rich, structurally complete responses.

This study conducts a systematic cross-lingual analysis of counseling dialogues across seven datasets in Chinese, Japanese, and English. The author introduces a two-stage identification pipeline: first, a rule-based filter identifies short candidate utterances; second, an LLM verifies these candidates within their specific dialogue context to confirm their interactional function. The study then evaluates how various models—ranging from general-purpose LLMs to counseling-specific models—perform when tasked with generating responses in contexts where human counselors would naturally use minimal responses.

• Data Disparity: Minimal responses are prevalent in human-collected datasets but are nearly absent in LLM-generated datasets. This suggests that synthetic training data may be inadvertently teaching models to avoid brief, interactional responses.
• Model Behavior: While strong commercial LLMs can generate minimal responses when explicitly instructed, they struggle to identify when such responses are appropriate under standard prompting. Models fine-tuned on synthetic data perform particularly poorly, consistently producing longer, more verbose responses.
• Evaluation Bias: Current LLM-based evaluation methods tend to undervalue minimal responses. Because these evaluators often reward comprehensiveness and informativeness, they penalize brief, interactionally appropriate responses, creating a feedback loop that discourages the development of more natural, listening-oriented counseling models.