Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RefuteBench: Evaluating Refuting Instruction-Following for Large Language Models

Jianhao Yan, Yun Luo, Yue Zhang

TL;DR

RefuteBench investigates how large language models respond to refuting feedback in multi-turn interactions, addressing a practical need to have models adapt to user corrections. It introduces a benchmark across three tasks—Question Answering, Machine Translation, and Email Writing—with single- and multi-feedback settings, and two core metrics, $FA$ and $RR$, to quantify feedback acceptance and follow-through. The study finds that many LLMs rely on internal knowledge and forget feedback as dialogue progresses, with GPT-4 and Claude-2 showing relatively higher flexibility than ChatGPT or Mistral. A lightweight recall-and-repeat prompting strategy is proposed and shown to substantially improve $RR$ across tasks, demonstrating a practical approach to enhance responsiveness without changing model parameters. The work releases data and code to spur further research into robust instruction-following under refuting feedback and highlights limitations and future directions in multi-round, real-world interactions.

Abstract

The application scope of large language models (LLMs) is increasingly expanding. In practical use, users might provide feedback based on the model's output, hoping for a responsive model that can complete responses according to their feedback. Whether the model can appropriately respond to users' refuting feedback and consistently follow through with execution has not been thoroughly analyzed. In light of this, this paper proposes a comprehensive benchmark, RefuteBench, covering tasks such as question answering, machine translation, and email writing. The evaluation aims to assess whether models can positively accept feedback in form of refuting instructions and whether they can consistently adhere to user demands throughout the conversation. We conduct evaluations on numerous LLMs and find that LLMs are stubborn, i.e. exhibit inclination to their internal knowledge, often failing to comply with user feedback. Additionally, as the length of the conversation increases, models gradually forget the user's stated feedback and roll back to their own responses. We further propose a recall-and-repeat prompts as a simple and effective way to enhance the model's responsiveness to feedback.

RefuteBench: Evaluating Refuting Instruction-Following for Large Language Models

TL;DR

RefuteBench investigates how large language models respond to refuting feedback in multi-turn interactions, addressing a practical need to have models adapt to user corrections. It introduces a benchmark across three tasks—Question Answering, Machine Translation, and Email Writing—with single- and multi-feedback settings, and two core metrics, and , to quantify feedback acceptance and follow-through. The study finds that many LLMs rely on internal knowledge and forget feedback as dialogue progresses, with GPT-4 and Claude-2 showing relatively higher flexibility than ChatGPT or Mistral. A lightweight recall-and-repeat prompting strategy is proposed and shown to substantially improve across tasks, demonstrating a practical approach to enhance responsiveness without changing model parameters. The work releases data and code to spur further research into robust instruction-following under refuting feedback and highlights limitations and future directions in multi-round, real-world interactions.

Abstract

The application scope of large language models (LLMs) is increasingly expanding. In practical use, users might provide feedback based on the model's output, hoping for a responsive model that can complete responses according to their feedback. Whether the model can appropriately respond to users' refuting feedback and consistently follow through with execution has not been thoroughly analyzed. In light of this, this paper proposes a comprehensive benchmark, RefuteBench, covering tasks such as question answering, machine translation, and email writing. The evaluation aims to assess whether models can positively accept feedback in form of refuting instructions and whether they can consistently adhere to user demands throughout the conversation. We conduct evaluations on numerous LLMs and find that LLMs are stubborn, i.e. exhibit inclination to their internal knowledge, often failing to comply with user feedback. Additionally, as the length of the conversation increases, models gradually forget the user's stated feedback and roll back to their own responses. We further propose a recall-and-repeat prompts as a simple and effective way to enhance the model's responsiveness to feedback.
Paper Structure (40 sections, 1 equation, 4 figures, 9 tables)

This paper contains 40 sections, 1 equation, 4 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Instruction Following
  4. Model Editing
  5. Retrieval-based Prompting Methods
  6. Problem Definition
  7. Single-Feedback
  8. Multi-Feedback
  9. Data Collection
  10. Knowledge -- Question Answering
  11. Multilinguality -- Machine Translation
  12. Writing -- Email
  13. Experimental Results
  14. Models
  15. Evaluation Metrics
  16. ...and 25 more sections

Figures (4)

  • Figure 1: Examples of Single-Feedback to illustrate our refuting dialogue process. The refuting instruction is given after LLMs' make their response to users' initial query.
  • Figure 2: The performance RR (Response Rate) and FA (Feedback Acceptance) of QA in different LLMs, which shows a positive correlation between RR and FA.
  • Figure 3: The performance RR (Response Rate) and FA (Feedback Acceptance) of QA in different LLMs, which shows a positive correlation between RR and FA.
  • Figure 4: Response rate against the number of context length.