CRWMTR: Causal Reasoning and Weighting Mechanisms of Thought Reasoning in Large Language Models

Large Language Models (LLMs) have profoundly improved the performance in natural language tasks and recently Chain of Thoughts (CoTs) combining Knowledge Graphs (KGs) has further improved reasoning. However, LLMs lack effective causality identification and weight adjustment mechanisms when dealing with complex problems, limiting the accuracy of multi-hop reasoning. To address the challenge, we propose the Causal Reasoning and Weighting Mechanisms of Thought Reasoning in Large Language Models (CRWMTR), which first brings the Key factor identification into complex multi-hop reasoning. For complex problem solving task, we recursively optimise the inference paths in the KGs by assigning weights <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$W$</tex> and scores <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$S$</tex> to the initial and candidate entities to guide the ‘merge’ and ‘delete’ operations. We propose a new self-correcting mechanism that can quantify and unify causal factors step by step during multi-hop reasoning, Experiments show CRWMTR outperform the state-of-the-art methods.