Counterfactual Debating with Preset Stances for Hallucination Elimination of LLMs

TL;DR

Addressing hallucination in LLMs, the paper identifies overconfidence in self-correction and diverse sampling as key weaknesses. It introduces Counterfactual Multi-Agent Debate (CFMAD), which presets LLM stances to generate abductions and uses adversarial critic debates evaluated by a judge to identify the correct answer. Through experiments on four datasets across three tasks (fact-checking, reading comprehension, and commonsense reasoning), CFMAD outperforms baselines including CoT, Self-Reflection, Self-Consistency, Self-Contrast, and MAD. The work demonstrates that counterfactual debates provide extra diagnostic cues and reduce reliance on biased initial answers, though it incurs higher computation and requires careful stance enumeration.

Abstract

Large Language Models (LLMs) excel in various natural language processing tasks but struggle with hallucination issues. Existing solutions have considered utilizing LLMs' inherent reasoning abilities to alleviate hallucination, such as self-correction and diverse sampling methods. However, these methods often overtrust LLMs' initial answers due to inherent biases. The key to alleviating this issue lies in overriding LLMs' inherent biases for answer inspection. To this end, we propose a CounterFactual Multi-Agent Debate (CFMAD) framework. CFMAD presets the stances of LLMs to override their inherent biases by compelling LLMs to generate justifications for a predetermined answer's correctness. The LLMs with different predetermined stances are engaged with a skeptical critic for counterfactual debate on the rationality of generated justifications. Finally, the debate process is evaluated by a third-party judge to determine the final answer. Extensive experiments on four datasets of three tasks demonstrate the superiority of CFMAD over existing methods.

Figures (6)

• Figure 1: Comparison of CFMAD with self-correction and diverse sampling methods. CFMAD presets stances for LLMs to override their inherent biases.
• Figure 2: Proportion of the overconfident answers among all incorrect answers.
• Figure 3: Illustration of CFMAD framework with two stages. In the abduction generation stage, we initialize multiple LLM agents, each configured to assume a specific answer is correct and to generate supporting abductions. In the subsequent counterfactual debate stage, each agent is challenged by a critical evaluator for debating. The debating processes are assessed by a third-party judge for final adjudication.
• Figure 4: Proportion of changes in initial stances. "Valid" means the stances changed from incorrect to correct. "Invalid" represents the stances changed from correct to incorrect.
• Figure 5: The final judgment on inconsistent stances. "Valid" means that the judge makes a correct judgment while "Invalid" denotes making an incorrect judgment.