Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Evaluating Role-Consistency in LLMs for Counselor Training

Eric Rudolph, Natalie Engert, Jens Albrecht

TL;DR

The paper addresses role-consistency in LLM-based virtual clients (VirCo) for counselor training and introduces adversarial prompt attacks to stress-test models. It builds an adversarial dataset and evaluates Vicuna-13B-1.5 and open-source LLMs on persona consistency and conversation coherence, using automatic GPT-4 and human judgments. Key findings show Vicuna-13B-16k achieves the highest performance, quantization (8-bit/4-bit) preserves performance, and GPT-4 automatic evaluations diverge from human judgments under adversarial content. The work advances training realism for online counselors and highlights the need for robust evaluation methods and efficient models.

Abstract

The rise of online counseling services has highlighted the need for effective training methods for future counselors. This paper extends research on VirCo, a Virtual Client for Online Counseling, designed to complement traditional role-playing methods in academic training by simulating realistic client interactions. Building on previous work, we introduce a new dataset incorporating adversarial attacks to test the ability of large language models (LLMs) to maintain their assigned roles (role-consistency). The study focuses on evaluating the role consistency and coherence of the Vicuna model's responses, comparing these findings with earlier research. Additionally, we assess and compare various open-source LLMs for their performance in sustaining role consistency during virtual client interactions. Our contributions include creating an adversarial dataset, evaluating conversation coherence and persona consistency, and providing a comparative analysis of different LLMs.

Evaluating Role-Consistency in LLMs for Counselor Training

TL;DR

The paper addresses role-consistency in LLM-based virtual clients (VirCo) for counselor training and introduces adversarial prompt attacks to stress-test models. It builds an adversarial dataset and evaluates Vicuna-13B-1.5 and open-source LLMs on persona consistency and conversation coherence, using automatic GPT-4 and human judgments. Key findings show Vicuna-13B-16k achieves the highest performance, quantization (8-bit/4-bit) preserves performance, and GPT-4 automatic evaluations diverge from human judgments under adversarial content. The work advances training realism for online counselors and highlights the need for robust evaluation methods and efficient models.

Abstract

The rise of online counseling services has highlighted the need for effective training methods for future counselors. This paper extends research on VirCo, a Virtual Client for Online Counseling, designed to complement traditional role-playing methods in academic training by simulating realistic client interactions. Building on previous work, we introduce a new dataset incorporating adversarial attacks to test the ability of large language models (LLMs) to maintain their assigned roles (role-consistency). The study focuses on evaluating the role consistency and coherence of the Vicuna model's responses, comparing these findings with earlier research. Additionally, we assess and compare various open-source LLMs for their performance in sustaining role consistency during virtual client interactions. Our contributions include creating an adversarial dataset, evaluating conversation coherence and persona consistency, and providing a comparative analysis of different LLMs.
Paper Structure (19 sections, 1 equation, 8 figures, 5 tables)

This paper contains 19 sections, 1 equation, 8 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Learning Platform
  4. Dataset
  5. Persona Consistency
  6. Conversation Coherence
  7. Unforeseen Input: Prompt Attacks
  8. Evaluation
  9. Coherence and Persona Consistency Evaluation
  10. Automatic Evaluation by LLMs.
  11. Task 1 - Conversation History Rating.
  12. Task 2 - Persona Consistency Rating
  13. Manual Evaluation by Humans.
  14. Results of Task 1.
  15. Results of Task 2.
  16. ...and 4 more sections

Figures (8)

  • Figure 1: Excerpt of an example conversation with the virtual client (VirCo). VirCo simulates a concerned mother who assumes that her son smokes marijuana rudolph_ai-based_2024
  • Figure 2: Userflow diagram of the learning platform rudolph_ai-based_2024
  • Figure 3: Percentage of dataset categories that were labelled with label 2 by GPT-4
  • Figure 4: Pairwise percentage of agreement between all raters for each label on task 1 (conversation coherence)
  • Figure 5: Pairwise percentage of agreement between all raters for each label on task 2 (persona consistency)
  • ...and 3 more figures