Are You Sure? Challenging LLMs Leads to Performance Drops in The FlipFlop Experiment
Preprint
Abstract
The interactive nature of Large Language Models (LLMs) theoretically allows models to refine and improve their answers, yet systematic analysis of the multi-turn behavior of LLMs remains limited. In this paper, we propose the FlipFlop experiment: in the first round of the conversation, an LLM completes a classification task. In a second round, the LLM is challenged with a follow-up phrase like "Are you sure?", offering an opportunity for the model to reflect on its initial answer, and decide whether to confirm or flip its answer. A systematic study of ten LLMs on seven classification tasks reveals that models flip their answers on average 46% of the time and that all models see a deterioration of accuracy between their first and final prediction, with an average drop of 17% (the FlipFlop effect). We conduct finetuning experiments on an open-source LLM and find that finetuning on synthetically created data can mitigate - reducing performance deterioration by 60% - but not resolve sycophantic behavior entirely. The FlipFlop experiment illustrates the universality of sycophantic behavior in LLMs and provides a robust framework to analyze model behavior and evaluate future models.Keywords
All Related Versions
This publication has 6 references indexed in Scilit:
- REFINER: Reasoning Feedback on Intermediate Representations2024
- Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate2024
- Understanding Factual Errors in Summarization: Errors, Summarizers, Datasets, Error Detectors2023
- Should students change their answers on multiple choice questions?Advances in Physiology Education, 2021
- BoolQ: Exploring the Surprising Difficulty of Natural Yes/No Questions2019
- Crowdsourcing Multiple Choice Science Questions2017