Abstract:With the widespread use of language models (LMs) in NLP tasks, researchers have discovered the potential of Chain-of-thought (CoT) to assist LMs in accomplishing complex reasoning tasks by generating intermediate steps. However, human thought processes are often non-linear, rather than simply sequential chains of thoughts. Therefore, we propose Graph-of-Thought (GoT) reasoning, which models human thought processes not only as a chain but also as a graph. By representing thought units as nodes and connections between them as edges, our approach captures the non-sequential nature of human thinking and allows for a more realistic modeling of thought processes. GoT adopts a two-stage framework with an additional GoT encoder for thought graph representation and fuses the graph representation with the original input representation through a gated fusion mechanism. We evaluate GoT's performance on a text-only reasoning task (AQUA-RAT) and a multimodal reasoning task (ScienceQA). Our model achieves significant improvement over the strong CoT baseline on the AQUA-RAT test set and boosts accuracy from 85.19% to 87.59% using the T5-base model over the state-of-the-art Multimodal-CoT on the ScienceQA test set.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that existing language models, when dealing with complex reasoning tasks, only rely on the linear "Chain - of - Thought" (CoT) method and overlook the non - linear and jumpy characteristics in the human thinking process. To overcome this limitation, the author proposes the "Graph - of - Thought" (GoT) method. By constructing a thought graph to simulate the human reasoning process, it can more realistically reflect the human thinking pattern. Specifically, the GoT method not only represents thinking units as nodes but also represents the connections between them as edges, which can capture the non - sequential nature of human thinking, enabling the model to perform more complex and realistic reasoning. ### Main Contributions: 1. **Proposing the GoT method**: Simulate the human reasoning process by constructing a thought graph to capture non - linear and jumpy thinking characteristics. 2. **Two - stage framework**: GoT adopts a two - stage framework. In the first stage, it generates intermediate reasoning steps (rationales), and in the second stage, it generates the final answer based on these reasoning steps. 3. **Multimodal reasoning**: Besides text features, GoT also integrates visual features and thought - graph features, which is suitable for multimodal reasoning tasks. 4. **Performance improvement**: On the AQUA - RAT and ScienceQA datasets, the GoT method significantly outperforms the existing CoT baseline models. Especially on the ScienceQA dataset, the performance of GoT even exceeds the human level. ### Specific Implementations: - **GoT Construction**: Use the Extract - Clustering - Coreference (ECC) process to extract and construct the thought graph from the input text. - **Encoding and Fusion**: Use the Graph Attention Network (GAT) to encode the thought graph and fuse the text, image, and thought - graph features together through a gated fusion mechanism. - **Experimental Verification**: Conducted extensive experiments on the AQUA - RAT and ScienceQA datasets to verify the effectiveness and superiority of the GoT method. ### Experimental Results: - **AQUA - RAT Dataset**: The ROUGE - L score of the GoT method in the reasoning stage has increased by 0.78, and the accuracy of the final answer has increased by 2.00%. - **ScienceQA Dataset**: The GoT method has achieved the best performance in all subjects and grades, with an overall accuracy rate of 87.59%, which is 2.40% higher than the existing Multimodal - CoT method. ### Conclusion: By introducing thought - graph features, the GoT method shows significant advantages in handling complex reasoning tasks, surpassing the traditional CoT method. This indicates that using two - dimensional thought graphs for reasoning is a more effective strategy and has broad application potential.

Beyond Chain-of-Thought, Effective Graph-of-Thought Reasoning in Language Models

Beyond Chain-of-Thought, Effective Graph-of-Thought Reasoning in Large Language Models

Graph Chain-of-Thought: Augmenting Large Language Models by Reasoning on Graphs

Understanding Reasoning in Chain-of-Thought from the Hopfieldian View

Multimodal Chain-of-Thought Reasoning in Language Models

Graph of Thoughts: Solving Elaborate Problems with Large Language Models

CoTAR: Chain-of-Thought Attribution Reasoning with Multi-level Granularity

Thinking Like an Expert:Multimodal Hypergraph-of-Thought (HoT) Reasoning to boost Foundation Modals

To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning

Supervised Chain of Thought

ChatCoT: Tool-Augmented Chain-of-Thought Reasoning on Chat-based Large Language Models

Markov Chain of Thought for Efficient Mathematical Reasoning

Visual Chain of Thought: Bridging Logical Gaps with Multimodal Infillings

A Hopfieldian View-based Interpretation for Chain-of-Thought Reasoning

Expediting and Elevating Large Language Model Reasoning via Hidden Chain-of-Thought Decoding

Think-on-Graph: Deep and Responsible Reasoning of Large Language Model with Knowledge Graph

Strategic Chain-of-Thought: Guiding Accurate Reasoning in LLMs through Strategy Elicitation

Improve Vision Language Model Chain-of-thought Reasoning

Rethinking Chain-of-Thought from the Perspective of Self-Training