# LLMs are Superior Feedback Providers: Bootstrapping Reasoning for Lie Detection with Self-Generated Feedback

**Authors:** Tanushree Banerjee, Richard Zhu, Runzhe Yang, Karthik Narasimhan  
**arXiv:** [2408.13915](https://arxiv.org/abs/2408.13915)  
**Date:** 2024-08-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.13915v1/

## Key points

- Novel bootstrapping framework using self-generated feedback to improve reasoning abilities of Large Language Models (LLMs) in detecting betrayal and deception in Diplomacy games
- Framework comprises three stages: suggestion, feedback collection, and modification
- LLM-generated feedback showed superior quality and significantly enhanced model's ability to detect lies, achieving a 39% improvement in lying-F1 score without additional training data
- LLM-generated feedback was longer and more informative compared to human feedback, outperforming human feedback by 29% in lying-F1 score while being more cost-effective
- Additional human study conducted to identify common errors made by LLMs like GPT-3 and consider ethics considerations
- Leveraging LLM-generated feedback can enhance model performance and offer an economical alternative for improving lie detection capabilities in natural language tasks

## Summary

In this study, we present a novel bootstrapping framework that utilizes self-generated feedback to improve the reasoning abilities of Large Language Models (LLMs) in detecting betrayal and deception in Diplomacy games. The framework comprises three stages: suggestion, feedback collection, and modification. Initially, a cost-effective language model generates predictions based on the game state and dialogue. Subsequently, a more advanced language model refines these predictions using auto-generated feedback. Our research compared the effectiveness of LLM-generated feedback with feedback from professional human players and found that LLM-generated feedback exhibited superior quality and significantly enhanced the model's ability to detect lies. This approach achieved an impressive 39% improvement in lying-F1 score without requiring additional training data, rivaling results from state-of-the-art supervised learning techniques. Furthermore, LLM-generated feedback was longer and provided more informative insights about potential missing predictions compared to human feedback. Notably, LLM-generated feedback outperformed human feedback by 29% in lying-F1 score while also being a more cost-effective solution. To better understand their limitations, we conducted an additional human study to identify common errors made by LLMs like GPT-3. We also took into account ethics considerations to ensure that studying deception did not have unintended consequences of improving deception tactics. Our findings suggest that leveraging LLM-generated feedback can enhance model performance and offer an economical alternative for improving lie detection capabilities in natural language tasks. However, it is important to note that OpenAI's GPT-4 model is not yet open-source, limiting access to advanced language models for research purposes. Despite this limitation, our study demonstrates the potential of utilizing LLMs for enhancing reasoning capabilities in detecting deception without major real-world consequences.