Random forest of thoughts: Reasoning path fusion for LLM inference in computational social science

Abstract

Large language models (LLMs) have demonstrated significant promise for reasoning problems. They are among the leading techniques for context inference, particularly in scenarios with strong sequential dependencies, where earlier inputs dynamically influence subsequent responses. However, existing reasoning paradigms such as X-of-thoughts (XoT) typically rely on unidirectional, left-to-right inference with limited inference paths. This renders them ineffective in handling inherent skip logic and multi-path reasoning, especially for contexts such as a multi-turn social survey. To address this, we propose Random Forest of Thoughts (RFoT), a novel prompting framework grounded in the principles of reasoning path fusion for skip logic. It uses Iterative Chain-of-Thought (ICoT) prompting to generate a diverse set of reasoning thoughts. These thoughts are then assessed using a cooperative contribution evaluator to estimate their contribution. By randomly sampling and fusing the top-$k$ reasoning thoughts, RFoT simulates uncertain skip logic and constructs a rich forest of plausible thoughts. This enables it to achieve robust multi-path reasoning, where each question sequence formed by the skip logic is treated as an independent reasoning path. RFoT is validated on two classic social problems featuring strong skip logic, using three open-source LLMs and five datasets that have been categorized as structured social surveys and public social media data. Experimental results demonstrate that RFoT significantly enhances inference performance on problems that require complex, non-linear reasoning across both survey and social media data. The transparency and trustworthiness of the results stem from the interpretable fusion of diverse reasoning paths and the principled integration of cooperative evaluation mechanisms.