SAGE: A Strategy-Aware Graph-Enhanced Generation Framework For Online Counseling

2026-04-29 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, AI for Mental Health · Depth: Expert, quick

Summary

SAGE (Strategy-Aware Graph-Enhanced) is a new framework designed to integrate structured clinical knowledge with generative AI for online mental health counseling. This framework addresses the complexity of theory-driven counseling, which often requires integrating psychological frameworks, distress signals, and strategic intervention planning, a capability frequently absent in general-purpose Large Language Models (LLMs). SAGE builds a heterogeneous graph that combines conversational dynamics with a psychologically grounded layer, linking interactions to a theory-driven lexicon. Its architecture uses a Next Strategy Classifier to determine the best therapeutic intervention, followed by a Graph-Aware Attention mechanism that projects graph-derived structural signals into soft prompts. This process conditions the LLM to produce clinically deep responses, outperforming baseline models in strategy prediction and response quality, as validated by automated metrics and expert human evaluation.

Key takeaway

For research scientists developing AI for sensitive domains like mental health, SAGE demonstrates a critical approach to embedding clinical reasoning into generative models. You should consider integrating structured knowledge graphs and explicit strategy classification to ensure AI outputs are not only coherent but also therapeutically sound and safe. This framework offers a blueprint for building decision-support tools that augment human expertise in high-stakes scenarios.

Key insights

SAGE integrates clinical knowledge graphs with LLMs to enhance strategy-aware, theory-driven online mental health counseling.

Principles

• Clinical reasoning requires simultaneous integration of frameworks and signals.
• Graph structures can unify conversational dynamics with psychological lexicons.

Method

SAGE employs a Next Strategy Classifier to identify optimal interventions, then uses a Graph-Aware Attention mechanism to project graph signals into soft prompts, conditioning an LLM for clinically deep responses.

In practice

• Use heterogeneous graphs to unify diverse data types.
• Condition LLMs with structural signals for domain-specific generation.

Topics

• SAGE Framework
• Online Counseling
• Large Language Models
• Heterogeneous Graphs
• Strategy Prediction

Best for: Research Scientist, AI Scientist, NLP Engineer, Domain Expert

Related on AIssential

• SAGE: Stochastic Prompt Optimization via Agent-Guided Exploration — Automatic prompt optimization is a black-box search problem, best approached by combining qualitative diagnosis with quantitative validation.
• Integrating Graphs, Large Language Models, and Agents: Reasoning and Retrieval — Integrating graphs with LLMs enhances reasoning, retrieval, and decision-making by combining structured knowledge with semantic understanding.
• Decoupled and Divergence-Conditioned Prompt for Multi-domain Dynamic Graph Foundation Models — DyGFM enables multi-domain dynamic graph foundation models by decoupling semantics and dynamics, and managing divergence.
• Connecting the Dots with Context Graphs — Stephen Chin, Neo4j — Context graphs integrate LLMs with knowledge graphs to provide grounded, explainable, and auditable AI-driven decision-making by connecting disparate enterprise data.
• SAGE: Stochastic Prompt Optimization via Agent-Guided Exploration — Automatic prompt optimization is effectively a black-box search problem, where agent-guided exploration can yield robust gains in complex dialogue systems.
• Beyond the Individual: Virtualizing Multi-Disciplinary Reasoning for Clinical Intake via Collaborative Agents — Aegle virtualizes MDT reasoning for clinical intake using a multi-agent framework, improving diagnostic accuracy and documentation quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.