[논문 리뷰] Causality for Large Language Models

Anpeng Wu, Kun Kuang외 8명|arXiv (Cornell University)|2024. 10. 20.

Topic Modeling인용 수 5

한 줄 요약

한 문장 직접 답변 요약: 종합적인 조사로 LLM 전반에 걸쳐 사전 훈련, 미세 조정, 정렬, 추론, 평가에 인과성을 통합해 그릇된 상관관계와 편향에 의존하는 문제를 극복한다.

ABSTRACT

Recent breakthroughs in artificial intelligence have driven a paradigm shift, where large language models (LLMs) with billions or trillions of parameters are trained on vast datasets, achieving unprecedented success across a series of language tasks. However, despite these successes, LLMs still rely on probabilistic modeling, which often captures spurious correlations rooted in linguistic patterns and social stereotypes, rather than the true causal relationships between entities and events. This limitation renders LLMs vulnerable to issues such as demographic biases, social stereotypes, and LLM hallucinations. These challenges highlight the urgent need to integrate causality into LLMs, moving beyond correlation-driven paradigms to build more reliable and ethically aligned AI systems. While many existing surveys and studies focus on utilizing prompt engineering to activate LLMs for causal knowledge or developing benchmarks to assess their causal reasoning abilities, most of these efforts rely on human intervention to activate pre-trained models. How to embed causality into the training process of LLMs and build more general and intelligent models remains unexplored. Recent research highlights that LLMs function as causal parrots, capable of reciting causal knowledge without truly understanding or applying it. These prompt-based methods are still limited to human interventional improvements. This survey aims to address this gap by exploring how causality can enhance LLMs at every stage of their lifecycle-from token embedding learning and foundation model training to fine-tuning, alignment, inference, and evaluation-paving the way for more interpretable, reliable, and causally-informed models. Additionally, we further outline six promising future directions to advance LLM development, enhance their causal reasoning capabilities, and address the current limitations these models face.

연구 동기 및 목표

• LLMs에 인과 추론의 통합을 촉진하여 상관관계 기반의 한계와 편향을 극복한다.
• 사전훈련, 미세조정, 정렬, 추론, 평가 단계에서 인과성을 어떻게 도입할 수 있는지 보여주는 라이프사이클 프레임워크를 제시한다.
• 인과성 기반 기법을 분류하고 LLM 개발 단계에 매핑한다.
• 도전과제를 강조하고 인과성 정보가 반영된 LLM의 향후 방향을 제시한다.

제안 방법

• 사전훈련, 미세조정, 정렬, 추론, 평가의 다섯 가지 LLM 단계에 걸친 인과성 영감 기법을 검토한다.
• 세 가지 인과적 사전훈련 접근법 제안: 편향 제거 토큰 임베딩, 반사실적(대응) 학습 코퍼스, 그리고 인과 기초 모델 프레임워크.
• 미세조정, 정렬, 추론에 대한 인과 강화 프로세스 개요: 인과 발견, 인과 효과, 그리고 반사실적 추론을 포함.

실험 결과

연구 질문

• RQ1 How can causality be embedded into the training and lifecycle of LLMs to surpass correlation-based learning?
• RQ2 What techniques at each stage (pre-training, fine-tuning, alignment, inference, evaluation) best promote causal reasoning in LLMs?
• RQ3 What are the main challenges and future directions for building causally-informed LLMs?

주요 결과

• Causality-based techniques across stages can address biases, reliability, and interpretability in LLMs.
• Current causal prompts rely on human intervention and prompts rather than intrinsic model understanding; embedding causality into training is essential.
• Six future directions are proposed to advance causal reasoning and overcome limitations of LLMs.