Source-linked AI summary
Psy-Copilot: Visual Chain of Thought for Counseling
Keqi Chen, Zekai Sun, Huijun Lian, Yingming Gao, Ya Li
TL;DR
Therapists need traceable, reliable support rather than client-facing black-box agents; this paper pairs Psy-COT with a retrieval-based counseling copilot. Psy-Agent outperformed baseline models across four reported dialogue metrics.
Problem
Many mental-support dialogue agents interact directly with clients, while therapists need a copilot that offers response candidates and traceable, reliable information.
Source: Introduction · p. 1Method
Psy-COT encodes 941 counseling sessions and extracted strategy and causal links in dialogue- and reasoning-specific indexes, which Psy-Copilot uses for dialogue-first, strategy-second retrieval.
Source: Introduction · p. 1, Construction of COT Graph · p. 2, Multi-Level Index for Retrieval · p. 2Results
Psy-Agent outperformed baseline models on Fluency, Helpfulness, Naturalness, and Comforting Effectiveness in the reported multi-turn dialogue evaluation.
Source: Objective Evaluation · p. 2Takeaways & Limitations
The authors present Psy-Copilot as a helpful assistant for psychotherapists and aim to foster AI–therapist synergy through visible response crafting.
Source: Conclusion · p. 2
Abstract
from arXiv · showhide
Large language models (LLMs) are becoming increasingly popular in the field of psychological counseling. However, when human therapists work with LLMs in therapy sessions, it is hard to understand how the model gives the answers. To address this, we have constructed Psy-COT, a graph designed to visualize the thought processes of LLMs during therapy sessions. The Psy-COT graph presents semi-structured counseling conversations alongside step-by-step annotations that capture the reasoning and insights of therapists. Moreover, we have developed Psy-Copilot, which is a conversational AI assistant designed to assist human psychological therapists in their consultations. It can offer traceable psycho-information based on retrieval, including response candidates, similar dialogue sessions, related strategies, and visual traces of results. We have also built an interactive platform for AI-assisted counseling. It has an interface that displays the relevant parts of the retrieval sub-graph. The Psy-Copilot is designed not to replace psychotherapists but to foster collaboration between AI and human therapists, thereby promoting mental health development. Our code and demo are both open-sourced and available for use.
Introduction
Existing support agents often address clients directly rather than collaborating with therapists, who need traceable, reliable information as well as response candidates. The paper introduces Psy-COT and Psy-Copilot to make counseling reasoning retrievable and visible within therapist assistance.
Introduction
- Psy-COT combines embedded dialogue sessions with a directed graph of events and strategies for visualizing and retrieving counseling reasoning.The introduction reports 941 embedded dialogue sessions.Source: Introduction · p. 1
- Psy-Copilot uses retrieval to generate traceable counseling responses and surface similar cases, relevant strategies, and visual traces of response generation.The retrieved information is intended to help therapists reason through consultations.Source: Introduction · p. 1
- Unlike knowledge graphs centered on entity-inclusion links, Psy-COT emphasizes event causality and the temporal evolution of counseling strategies.Its graph maps events and strategies to dialogue units.Source: Introduction · p. 1, Figure 2 · p. 1
- The project combines counseling-data collection, an explainable AI agent, and an interactive platform for therapist collaboration.The platform is intended to make outcomes visualized and comprehensible to human therapists.Source: Introduction · p. 1
Psy-COT Graph Construction
Psy-COT is built from counseling sessions and explanatory material, with extracted reasoning linked to dialogue and separate indexes for dialogue and chain-of-thought content.
Data Source
- The dataset comprises 941 psycho-counseling sessions collected from psychological blogs, each with dialogue and corresponding explanatory text.Source: Data Source · p. 2
Construction of COT Graph
- A multi-agent framework extracts strategy descriptions and causal relationships, then uses a sliding window to map chain-of-thought nodes to dialogue nodes.Strategy types and causal relationships are defined based on counseling research.Source: Construction of COT Graph · p. 2
Multi-Level Index for Retrieval
- Psy-COT uses separate indexes for dialogue and reasoning content because counseling explanations and dialogue differ in granularity and language.The explanations use specialized psychological terms, while dialogue is more euphemistic.Source: Multi-Level Index for Retrieval · p. 2
Psy-Copilot Agent
Psy-Copilot retrieves similar counseling dialogues and relevant strategies to generate responses, with related subgraphs displayed to therapists. Its multi-turn dialogue evaluation reports performance above baseline models across four metrics.
Retrieval Argument Generation
- Conversation history retrieves similar dialogues, then generated reasoning guides strategy retrieval; filtered dialogue and reasoning nodes inform response generation.Dialogue nodes become few-shot examples, while reasoning content supplies instructions.Source: Retrieval Argument Generation · p. 2
Demo UI
- The interactive platform dynamically displays related subgraphs during each dialogue round to help therapists trace the generation logic.Open-source Qwen-2.5-7B and Deepseek-V2 power the retrieval and chat agents.Source: Demo UI · p. 2, Base LLMs · p. 2
Objective Evaluation
- Because interactive multi-turn evaluation is time-consuming and difficult to quantify, GLM4-9b evaluates Psy-Copilot’s emotional intelligence.The reported assessment uses Fluency, Helpfulness, Naturalness, and Comforting Effectiveness.Source: Objective Evaluation · p. 2
- Psy-Agent outperforms baseline models across all four evaluation metrics.Table 1 covers Fluency, Helpfulness, Naturalness, and Comforting Effectiveness.Source: Table 1 · p. 2, Objective Evaluation · p. 2
Conclusion
The paper presents Psy-COT and Psy-Copilot as visual and conversational tools for traceable psychological counseling. It provides an open-source graph-construction pipeline and an interactive counseling platform intended to support collaboration between AI and therapists.
Conclusion
- Psy-COT and Psy-Copilot pair a visual graph with a conversational agent for traceable psychological counseling.The authors open-source the graph-construction pipeline and deploy an interactive platform, aiming to support AI–therapist synergy.Source: Conclusion · p. 2