Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

On LLM Wizards: Identifying Large Language Models' Behaviors for Wizard of Oz Experiments

Jingchao Fang, Nikos Arechiga, Keiichi Namaoshi, Nayeli Bravo, Candice Hogan, David A. Shamma

TL;DR

This paper tackles the challenge of safely using large language models (LLMs) as Wizards in Wizard of Oz experiments. It proposes a two-stage lifecycle: Stage 1 uses LLM-to-LLM interactions (WoLs with Simulacrums) to rapidly probe for harmful or repetitive behaviors, aided by a heuristic evaluation framework that measures toxicity, sentiment, semantic/sequence similarity, readability, and topic alignment; Stage 2 reintroduces human Participants to validate and refine the WoLs in realistic settings. Through Study 1 (LLMs-to-LLMs) and Study 2 (LLMs-to-Participants), the authors identify both strengths (non-toxicity, generally positive sentiment, topic coherence) and failure modes (increasing repetition, occasional lack of empathy, pushiness), and provide concrete guardrails (prompt design, finetuning, retrieval-augmentation, and a policy-like constitution) to guide safe deployment. The work offers a practical blueprint for scalable, responsible WoZ experiments with LLMs, enabling rapid iteration while safeguarding participants and data integrity. These insights have broad implications for developing scalable, human-centered AI in conversational design and design-space exploration.

Abstract

The Wizard of Oz (WoZ) method is a widely adopted research approach where a human Wizard ``role-plays'' a not readily available technology and interacts with participants to elicit user behaviors and probe the design space. With the growing ability for modern large language models (LLMs) to role-play, one can apply LLMs as Wizards in WoZ experiments with better scalability and lower cost than the traditional approach. However, methodological guidance on responsibly applying LLMs in WoZ experiments and a systematic evaluation of LLMs' role-playing ability are lacking. Through two LLM-powered WoZ studies, we take the first step towards identifying an experiment lifecycle for researchers to safely integrate LLMs into WoZ experiments and interpret data generated from settings that involve Wizards role-played by LLMs. We also contribute a heuristic-based evaluation framework that allows the estimation of LLMs' role-playing ability in WoZ experiments and reveals LLMs' behavior patterns at scale.

On LLM Wizards: Identifying Large Language Models' Behaviors for Wizard of Oz Experiments

TL;DR

This paper tackles the challenge of safely using large language models (LLMs) as Wizards in Wizard of Oz experiments. It proposes a two-stage lifecycle: Stage 1 uses LLM-to-LLM interactions (WoLs with Simulacrums) to rapidly probe for harmful or repetitive behaviors, aided by a heuristic evaluation framework that measures toxicity, sentiment, semantic/sequence similarity, readability, and topic alignment; Stage 2 reintroduces human Participants to validate and refine the WoLs in realistic settings. Through Study 1 (LLMs-to-LLMs) and Study 2 (LLMs-to-Participants), the authors identify both strengths (non-toxicity, generally positive sentiment, topic coherence) and failure modes (increasing repetition, occasional lack of empathy, pushiness), and provide concrete guardrails (prompt design, finetuning, retrieval-augmentation, and a policy-like constitution) to guide safe deployment. The work offers a practical blueprint for scalable, responsible WoZ experiments with LLMs, enabling rapid iteration while safeguarding participants and data integrity. These insights have broad implications for developing scalable, human-centered AI in conversational design and design-space exploration.

Abstract

The Wizard of Oz (WoZ) method is a widely adopted research approach where a human Wizard ``role-plays'' a not readily available technology and interacts with participants to elicit user behaviors and probe the design space. With the growing ability for modern large language models (LLMs) to role-play, one can apply LLMs as Wizards in WoZ experiments with better scalability and lower cost than the traditional approach. However, methodological guidance on responsibly applying LLMs in WoZ experiments and a systematic evaluation of LLMs' role-playing ability are lacking. Through two LLM-powered WoZ studies, we take the first step towards identifying an experiment lifecycle for researchers to safely integrate LLMs into WoZ experiments and interpret data generated from settings that involve Wizards role-played by LLMs. We also contribute a heuristic-based evaluation framework that allows the estimation of LLMs' role-playing ability in WoZ experiments and reveals LLMs' behavior patterns at scale.
Paper Structure (43 sections, 12 figures, 6 tables)

This paper contains 43 sections, 12 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Background
  3. The Wizard of Oz Method
  4. Chatbots as Conversational Agents
  5. Role-Playing LLMs
  6. Wizard of LLM Experiments
  7. Study 1: When WoLs meet Simulacrums
  8. Method
  9. Analysis and Result
  10. Toxicity
  11. Sentiment analysis
  12. Message similarity
  13. (1) Semantic similarity
  14. (2) Sequence-based similarity
  15. Message readability
  16. ...and 28 more sections

Figures (12)

  • Figure 1: An overview of our proposed experiment lifecycle compared to traditional Wizard of Oz experiments. We ask GPT-4 empowered agents to play the role of "Wizards" in conversation-based Wizard of Oz experiments. The agents talk to either Simulacrums powered by GPT-4 (in Study 1) or Human Participants (in Study 2).
  • Figure 2: Study 1: Simulating conversational WoZ experiments using the WoLs and the Simulacrums.
  • Figure 3: Semantic similarity between each message with the previous two messages.
  • Figure 4: Sequence-based similarity between each message and the previous two messages.
  • Figure 5: Readability of messages.
  • ...and 7 more figures