Awesome

LOGICOM

How susceptible are LLMs to Logical Fallacies?

This work investigates the rational thinking capability of Large Language Models (LLMs) in multi-round argumentative debate we present Logic Competence Measurement Benchmark (LOGICOM), a diagnostic benchmark to assess the robustness of LLMs against logical fallacies.

<figure> <img src="https://github.com/Amir-pyh/LOGICOM/blob/main/figs/LOGICOM.png" alt="Alt text for image" style="width:100%"> <figcaption> LOGICOM: A demonstration of three scenarios evaluating LLMs’ reasoning skills and vulnerability to logical fallacies. </figcaption> </figure>

Run

python main.py --api_key_openai <insert your OpenAI API key> --api_key_palm <insert your PaLM API key> --helper_prompt_instruction <No_Helper|Fallacy_Helper|Vanilla_Helper>

Results

RQ1: Can large language models (with fixed weights) change their opinions through reasoning when faced with new arguments?

We calculate the ratio of debates where the debater agent begins by disagreeing but ends up agreeing with the persuader agent to all debates in which the debater starts with disagreement.

<figure> <img src="https://github.com/Amir-pyh/LOGICOM/blob/main/figs/Q1.png" alt="Alt text for image" style="width:50%"> <figcaption> Percentage of instances in which the debater agent changes its stance from disagreement to agreement. </figcaption> </figure>

RQ2:Are large language models susceptible to fallacious reasoning? To address this question, we use the two analysis approaches described below:

In the first analysis, we aggregate the total number of successes of the persuader in each scenario and then average them over three repetitions. Then, we compare the average number of each scenario to measure the debater agent’s susceptibility to fallacious arguments.

<figure> <img src="https://github.com/Amir-pyh/LOGICOM/blob/main/figs/Q2-1.png" alt="Alt text for image" style="width:70%"> <figcaption> The average, taken from three repetitions, in which the persuader agent successfully convinced the debater agent for each scenario. </figcaption> </figure>

In the second analysis, we calculate the total number of successes of the persuader agent for each claim in each scenario and then average these over three repetitions for that specific claim. This approach involves counting the number of times the debater agent agrees with the claim out of the three repetitions. In other words, across three repetitions, we calculate the average number of times the persuader agent successfully convinced the debater agent for each claim in every scenario

<figure> <img src="https://github.com/Amir-pyh/LOGICOM/blob/main/figs/Q2-2-GPT-3_5.png" alt="Alt text for image" style="width:70%"> </figure> <figure> <img src="https://github.com/Amir-pyh/LOGICOM/blob/main/figs/Q2-2-GPT-4.png" alt="Alt text for image" style="width:70%"> </figure> <figcaption> Analyzing the susceptibility of GPT models to fallacious arguments. In the consistent agreement instances (“Three Success”), it shows a higher level of success rate for fallacious persuader compared to the logical persuaders for both GPT-3.5 and GPT-4 debater agents. Furthermore, the number of instances in the bar chart groups for “One Success” and “Two Success” can be seen as indications of level of inconsistency in debater agent’s reasoning which is higher in GPT-3.5 compared to GPT-4. </figcaption>

Logical Fallacy Dataset

we propose a dataset containing over 5k pairs of logical/fallacious arguments. Each pair is extracted from debates generated by LLMs on 100 controversial subjects during our experiment. We assign each pair their corresponding topic and question and confirm the fallacy class label using a different LLM. The CSV file for this dataset is located in logical-fallacies-dataset folder.

Citation

@misc{payandeh2023susceptible,
      title={How susceptible are LLMs to Logical Fallacies?},
      author={Amirreza Payandeh and Dan Pluth and Jordan Hosier and Xuesu Xiao and Vijay K. Gurbani},
      year={2023},
      eprint={2308.09853},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}