Reproduce “Prompting Large Language Models for Counterfactual Generation: An Empirical Study” on OpenResearch

About this paper

Yongqi Li, Mayi Xu, Xin Miao, Shen Zhou, Tieyun Qian

Researchers from Wuhan University systematically evaluated Large Language Models' (LLMs) ability to generate counterfactuals across various Natural Language Understanding (NLU) tasks. Their study reveals that LLM-generated counterfactuals effectively augment smaller language models for simpler tasks like sentiment analysis, but face limitations with complex tasks such as relation extraction, often due to issues with semantic constraints and inherent biases.

Key insights

• LLMs, particularly those with alignment techniques like instruction-tuning and RLHF, effectively generate counterfactuals for simpler NLU tasks (SA, NLI), often outperforming specialized internal knowledge methods.
• Increasing LLM parameter size alone does not proportionally improve counterfactual generation quality, challenging conventional scaling laws in this domain; however, alignment techniques significantly enhance performance.
• LLMs struggle with complex NLU tasks (NER, RE) due to issues like failure to consider entity constraints and inherent selection biases, which can lead to performance degradation in augmented models.
• Detailed task guidelines in prompts are crucial, but Chain-of-Thought prompting can be detrimental due to inconsistencies between LLM reasoning and final output.

Summary by alphaXiv.