Think More, Hallucinate Less: Mitigating Hallucinations via Dual Process of Fast and Slow Thinking

2025-01-02

AI Summary

To address pervasive factual hallucinations in large language model (LLM) text generation, this paper proposes HaluSearch: a framework that models text generation as a stepwise reasoning process, integrating Monte Carlo Tree Search (MCTS) with self-supervised reward modeling to establish a joint “tree search + self-evaluation” optimization paradigm. It is the first work to incorporate cognitive science’s dual-process theory into LLM inference mechanisms, introducing a hierarchical fast/slow thinking dynamic switching strategy that enables instance-level and step-level multi-granularity reasoning control. Evaluated on Chinese and English hallucination benchmarks, HaluSearch significantly outperforms state-of-the-art baselines, achieving substantial gains in factual accuracy while maintaining high reasoning quality and computational efficiency.

Technology Category

• Search and Optimization: Learning to Search
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Large language models (LLMs) demonstrate exceptional capabilities, yet still face the hallucination issue. Typical text generation approaches adopt an auto-regressive generation without deliberate reasoning, which often results in untrustworthy and factually inaccurate responses. In this paper, we propose HaluSearch, a novel framework that incorporates tree search-based algorithms (e.g. MCTS) to enable an explicit slow thinking generation process for mitigating hallucinations of LLMs during inference. Specifically, HaluSearch frames text generation as a step-by-step reasoning process, using a self-evaluation reward model to score each generation step and guide the tree search towards the most reliable generation pathway for fully exploiting the internal knowledge of LLMs. To balance efficiency and quality, we introduce a hierarchical thinking system switch mechanism inspired by the dual process theory in cognitive science, which dynamically alternates between fast and slow thinking modes at both the instance and step levels, adapting to the complexity of questions and reasoning states. We conduct extensive experiments on both English and Chinese datasets and the results show that our approach significantly outperforms baseline approaches.

Problem

• Large Language Models
• Factuality Improvement
• Accuracy Enhancement

Innovation

• HaluSearch
• Fast-and-Slow-Thinking
• Enhanced-Accuracy