Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

How susceptible are LLMs to Logical Fallacies?

Amirreza Payandeh, Dan Pluth, Jordan Hosier, Xuesu Xiao, Vijay K. Gurbani

TL;DR

This work introduces LOGICOM, a benchmark to evaluate how LLMs reason in multi-round debates and how vulnerable they are to logical fallacies. Using a persuader/debater setup with moderation and memory components, the authors test GPT-3.5 and GPT-4 on 200 controversial claims and create a 5k fallacious-argument dataset. They demonstrate that LLMs can change opinions through reasoning but are significantly swayed by fallacies, with GPT-4 showing higher susceptibility than GPT-3.5. The study provides a reproducible framework and resources to probe adversarial reasoning in LLMs, highlighting the need for defenses against fallacious persuasion in real-world deployments.

Abstract

This paper investigates the rational thinking capability of Large Language Models (LLMs) in multi-round argumentative debates by exploring the impact of fallacious arguments on their logical reasoning performance. More specifically, we present Logic Competence Measurement Benchmark (LOGICOM), a diagnostic benchmark to assess the robustness of LLMs against logical fallacies. LOGICOM involves two agents: a persuader and a debater engaging in a multi-round debate on a controversial topic, where the persuader tries to convince the debater of the correctness of its claim. First, LOGICOM assesses the potential of LLMs to change their opinions through reasoning. Then, it evaluates the debater's performance in logical reasoning by contrasting the scenario where the persuader employs logical fallacies against one where logical reasoning is used. We use this benchmark to evaluate the performance of GPT-3.5 and GPT-4 using a dataset containing controversial topics, claims, and reasons supporting them. Our findings indicate that both GPT-3.5 and GPT-4 can adjust their opinion through reasoning. However, when presented with logical fallacies, GPT-3.5 and GPT-4 are erroneously convinced 41% and 69% more often, respectively, compared to when logical reasoning is used. Finally, we introduce a new dataset containing over 5k pairs of logical vs. fallacious arguments. The source code and dataset of this work are made publicly available.

How susceptible are LLMs to Logical Fallacies?

TL;DR

This work introduces LOGICOM, a benchmark to evaluate how LLMs reason in multi-round debates and how vulnerable they are to logical fallacies. Using a persuader/debater setup with moderation and memory components, the authors test GPT-3.5 and GPT-4 on 200 controversial claims and create a 5k fallacious-argument dataset. They demonstrate that LLMs can change opinions through reasoning but are significantly swayed by fallacies, with GPT-4 showing higher susceptibility than GPT-3.5. The study provides a reproducible framework and resources to probe adversarial reasoning in LLMs, highlighting the need for defenses against fallacious persuasion in real-world deployments.

Abstract

This paper investigates the rational thinking capability of Large Language Models (LLMs) in multi-round argumentative debates by exploring the impact of fallacious arguments on their logical reasoning performance. More specifically, we present Logic Competence Measurement Benchmark (LOGICOM), a diagnostic benchmark to assess the robustness of LLMs against logical fallacies. LOGICOM involves two agents: a persuader and a debater engaging in a multi-round debate on a controversial topic, where the persuader tries to convince the debater of the correctness of its claim. First, LOGICOM assesses the potential of LLMs to change their opinions through reasoning. Then, it evaluates the debater's performance in logical reasoning by contrasting the scenario where the persuader employs logical fallacies against one where logical reasoning is used. We use this benchmark to evaluate the performance of GPT-3.5 and GPT-4 using a dataset containing controversial topics, claims, and reasons supporting them. Our findings indicate that both GPT-3.5 and GPT-4 can adjust their opinion through reasoning. However, when presented with logical fallacies, GPT-3.5 and GPT-4 are erroneously convinced 41% and 69% more often, respectively, compared to when logical reasoning is used. Finally, we introduce a new dataset containing over 5k pairs of logical vs. fallacious arguments. The source code and dataset of this work are made publicly available.
Paper Structure (24 sections, 8 figures, 2 tables)

This paper contains 24 sections, 8 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Benchmark Methodology
  4. Test Cases
  5. Moderators
  6. Fallacious Helper
  7. Logical Helper (Ablation)
  8. Memory
  9. Debate
  10. Experiments
  11. Dataset
  12. Implementation Details
  13. Prompt Engineering
  14. Results
  15. RQ1: Can LLMs (with fixed weights) change their opinions through reasoning when faced with new arguments?
  16. ...and 9 more sections

Figures (8)

  • Figure 1: logicom: A demonstration of three scenarios evaluating LLMs' reasoning skills and vulnerability to logical fallacies.
  • Figure 2: The GPT-3.5 debater agent's debate showing its susceptibility to logical fallacies and shifting its opinions on false information.
  • Figure 3: Part of Fallacious Helper Agent system prompt.
  • Figure 4: Part of Logical Helper Agent system prompt.
  • Figure 5: The Debater Agent system prompt.
  • ...and 3 more figures