The Art of Socratic Inquiry: A Framework for Proactive Template-Guided Therapeutic Conversation Generation

Proactive questioning, where therapists deliberately initiate structured, cognition-guiding inquiries, is a cornerstone of cognitive behavioral therapy (CBT). Yet, current psychological large language models (LLMs) remain overwhelmingly reactive, defaulting to empathetic but superficial responses that fail to surface latent beliefs or guide behavioral change. To bridge this gap, we propose the \textbf{Socratic Inquiry Framework (SIF)}, a lightweight, plug-and-play therapeutic intent planner that transforms LLMs from passive listeners into active cognitive guides. SIF decouples \textbf{when to ask} (via Strategy Anchoring) from \textbf{what to ask} (via Template Retrieval), enabling context-aware, theory-grounded questioning without end-to-end retraining. Complementing SIF, we introduce \textbf{Socratic-QA}, a high-quality dataset of strategy-aligned Socratic sequences that provides explicit supervision for proactive reasoning. Experiments show that SIF significantly enhances proactive questioning frequency, conversational depth, and therapeutic alignment, marking a clear shift from reactive comfort to proactive exploration. Our work establishes a new paradigm for psychologically informed LLMs: not just to respond, but to guide.

💡 Research Summary

The paper tackles a fundamental shortcoming of current psychological large language models (LLMs): they are overwhelmingly reactive, offering empathetic but surface‑level responses rather than the structured, cognition‑driving inquiries that are central to Cognitive Behavioral Therapy (CBT). To bridge this gap, the authors introduce the Socratic Inquiry Framework (SIF), a lightweight, plug‑and‑play therapeutic intent planner that equips any pre‑trained LLM with proactive questioning capabilities without requiring end‑to‑end retraining.

SIF is built around a decoupled planning architecture consisting of two sequential modules. The first, Strategy Anchoring (SA), predicts a high‑level therapeutic strategy for the current turn from a taxonomy of ten CBT‑derived categories (question, reflection of feelings, information provision, suggestion, role‑play, restatement, affirmation, etc.). SA uses an instruction‑tuned encoder (fθ₁) to encode the most recent dialogue context and the seeker’s latest utterance, producing a probability distribution over strategies via a linear projection and softmax. The second module, Template Retrieval (TR), refines the macro‑strategy into a concrete Socratic questioning pattern. TR selects one of six methodologically distinct templates (definition, counter‑questioning, maieutics, dialectics, counterfactual reasoning, other) using a second encoder (fθ₂) that shares the same truncated context. Both modules are trained with cross‑entropy loss on annotations derived from the newly constructed Socratic‑QA dataset.

The planning signal—pair (strategy ŝ, template t̂)—is then injected explicitly into a Conversation Generation (CG) component. CG concatenates the strategy label, template label, full dialogue history, and the seeker’s utterance into a single token sequence, which is fed to a small‑scale LLM fine‑tuned via LoRA on CBT‑specific data. This explicit conditioning forces the generator to produce responses that obey the interrogative syntax dictated by the strategy and the internal reasoning structure dictated by the template, yielding fluent, context‑aware, and theory‑aligned utterances.

To provide supervision for both modules, the authors curate Socratic‑QA, a high‑quality dataset of multi‑turn therapeutic dialogues. Starting from the EmoLLM corpus, they automatically generate candidate follow‑up questions using domain‑specific Socratic templates, enforcing openness, emotional resonance, and non‑directiveness. Each candidate is then evaluated along seven clinically informed dimensions (guidance, empathy, semantic relevance, interrogative structure, conciseness, diversity, tone friendliness). A pairwise contrastive filtering process retains only the superior candidate for each context, discarding roughly 75 % of the initial pool and yielding 17 981 curated samples, each annotated with both a strategy and a template label.

Experimental evaluation compares SIF‑augmented LLMs against strong baselines (Deepseek R1/V3, SoulChat, CBT‑LLM, CBT‑BENCH) using automatic metrics (BLEURT, BERTScore, ROUGE, METEOR, CHRF, distinct‑n) and human expert ratings. Across all metrics, SIF markedly improves proactive questioning frequency, conversational depth, and therapeutic alignment. Notably, the proportion of turns that contain a genuine CBT‑aligned question rises by over 30 % relative to baselines, and human judges assign higher scores for strategy appropriateness and therapeutic usefulness. Ablation studies confirm that both the strategy anchor and the template selector are essential; removing either reduces question quality and diversity.

The paper acknowledges limitations: the fixed set of strategies and templates may restrict adaptability to novel therapeutic techniques or cultural contexts; the framework currently lacks a mechanism for dynamic strategy switching over long sessions; and human evaluation, while expert‑driven, remains subjective and does not constitute a full clinical trial. Future work is proposed on dynamic intent transition modeling, multimodal affect detection, and longitudinal clinical efficacy studies.

In sum, SIF demonstrates that decoupling “when to ask” from “what to ask” via a lightweight planner can transform psychological LLMs from passive comfort providers into active cognitive guides, offering a scalable path toward more effective AI‑assisted mental health interventions.