Between Sessions, Not Instead of Sessions: LLM-Generated Check-Ins, Homework, and Reflection Prompts for Counseling Continuity

Abstract

This paper studies counseling continuity outside the therapy hour as the generation of three concrete between-session artifacts: a check-in, a homework task, and a reflection prompt. We frame the problem as a planning-and-realization layer for LLM-delivered follow-ups that remain grounded in the prior session rather than replacing clinical contact. Direct programmatic access to the originally targeted session-annotated counseling corpus was not available in the present reproducibility environment, so we built ContinuityBench-16,870, a proxy benchmark that preserves the task-critical session fields of summary, guide, stage, and reasoning. The benchmark contains 12,800 training examples, 1,600 development examples, 1,600 in-distribution test examples, and 870 compositional out-of-distribution test examples; it covers 15 concern categories, 16 coping strategies, 4 therapy stages, and 7.94% elevated-risk cases. We compare a stage-conditioned template baseline, TF-IDF retrieval, a Multinomial Naive Bayes planner, a Linear SVC planner, and a reasoning-aware multi-task planner (RMTP). On the full test set, RMTP reaches package ROUGE-L of 0.765, package BLEU of 0.699, exact plan accuracy of 0.695, risk recall of 1.000, and a safety-violation rate of 0.000. Relative to the strongest non-reasoning baseline, RMTP improves package ROUGE-L by 0.051, package BLEU by 0.062, and exact plan accuracy by 0.023. Homework generation remains the hardest artifact, while reflection prompts are easiest. Tone selection is the main residual error source. The results show that between-session support can be generated as a structured, safety-routed continuity layer and should be deployed between sessions, not instead of sessions.