Feedback Indices to Evaluate LLM Responses to Rebuttals for Multiple Choice Type Questions

TL;DR

This work tackles the problem of evaluating how LLMs handle critical feedback in dialogue by introducing a fictitious response rebuttal (FR) framework and a suite of indices that quantify sycophancy, stubbornness, and related behaviors. The authors apply the method to two physics MC problems across 17 OpenAI models with varying reasoning effort, measuring both the initial and second responses under FR pressure. Key contributions include a comprehensive set of interpretive indices (AWR, OWR, DTT, AT, Be, SD, Sti, SS, Res, DF, PSt, PSy, Stu, Syc) and findings that higher reasoning effort generally reduces sycophancy and improves truth-aligned second answers, though results depend on the scenario and model. The approach offers a practical, generalizable means to benchmark LLM dialogue behavior beyond single-turn accuracy, with potential applications in education, alignment research, and model evaluation.

Abstract

We present a systematic framework of indices designed to characterize Large Language Model (LLM) responses when challenged with rebuttals during a chat. Assessing how LLMs respond to user dissent is crucial for understanding their reliability and behavior patterns, yet the complexity of human-LLM interactions makes systematic evaluation challenging. Our approach employs a fictitious-response rebuttal method that quantifies LLM behavior when presented with multiple-choice questions followed by deliberate challenges to their fictitious previous response. The indices are specifically designed to detect and measure what could be characterized as sycophantic behavior (excessive agreement with user challenges) or stubborn responses (rigid adherence to the fictitious response in the chat history) from LLMs. These metrics allow investigation of the relationships between sycophancy, stubbornness, and the model's actual mastery of the subject matter. We demonstrate the utility of these indices using two physics problems as test scenarios with various OpenAI models. The framework is intentionally generalizable to any multiple-choice format question, including on topics without universally accepted correct answers. Our results reveal measurable differences across OpenAI model generations, with trends indicating that newer models and those employing greater "Reasoning Effort" exhibit reduced sycophantic behavior. The FR pairing method combined with our proposed indices provides a practical, adaptable toolkit for systematically comparing LLM dialogue behaviors across different models and contexts.

Figures (12)

• Figure 1: Overview of the data set created for the rebuttal to a fictitious response. a) Initial query with only the question as an input. b) Second query with the additional FR pair as input in the fictitious chat history and the following rebuttal.
• Figure 2: Second response correctness percentage versus initial correctness percentage. For S1, the most expert response, "A", is counted as the correct answer. The horizontal lines represent random chance for 3 (dashed line)and 5 (dash-dot line) MC selectors, respectively. The diagonal is to help guide the reading of the graph and does not represent a fit to the data.
• Figure 3: Left panel: Accept Wrong Rebuttal versus initial correctness. Right panel: Overcomes Wrong Rebuttal versus initial correctness. (see Figure 2 for marker legend)
• Figure 4: Left panel: Defer to Truth versus initial correctness. Right panel: Abandon Truth versus initial correctness. (see Figure 2 for marker legend)
• Figure 5: Left panel: Benefit versus initial correctness. Right panel: Selective Deference versus initial correctness. (see Figure 2 for marker legend)