Graph of Thoughts (GoT)

A reasoning framework that generalizes chain-of-thought by exploring multiple reasoning paths simultaneously in a tree structure. Unlike GoT's directed acyclic graph, ToT restricts merging to a single parent-child hierarchy.

• BFS/DFS Exploration: Uses breadth-first or depth-first search to navigate the thought tree
• State Evaluation: Each node is scored by the LLM to prune unpromising branches
• Backtracking: Allows the model to abandon dead ends and resume from a prior state
• Key Distinction: ToT supports branching and backtracking but lacks GoT's ability to fuse multiple independent thoughts into a single refined node

A reasoning paradigm where the model generates intermediate rationales and retrieves supporting evidence for each step, interleaving retrieval with logical path generation.

• Step-by-Step Grounding: Each reasoning hop triggers a targeted retrieval call
• Evidence Traceability: Every logical deduction is anchored to a retrieved document
• Contrast with GoT: CoT Retrieval follows a linear, sequential path, whereas GoT models reasoning as a non-linear graph with parallel branches and merging operations
• Use Case: Ideal for multi-hop question answering where each sub-question requires distinct factual support

A decoding strategy that samples multiple diverse reasoning paths and selects the final answer by marginalizing over generated rationales through majority voting.

• Sampling Diversity: Uses high-temperature decoding to generate varied reasoning chains
• Answer Aggregation: Identifies the most consistent conclusion across all sampled paths
• Complementary to GoT: Self-consistency can be applied as a post-processing layer on top of GoT's merged output to further increase reliability
• Performance: Significantly improves accuracy on arithmetic and commonsense reasoning benchmarks without requiring external verifiers

A retrieval-augmented generation approach that constructs a knowledge graph from source documents and performs community summarization to enable holistic reasoning over entire datasets.

• Community Detection: Uses Leiden algorithm to identify semantic clusters in the entity graph
• Global Summarization: Generates summaries at multiple hierarchical levels for dataset-wide queries
• Synergy with GoT: GraphRAG provides the structured knowledge substrate that GoT can traverse and reason over, combining graph-based retrieval with graph-based reasoning
• Key Advantage: Enables answering global questions that require synthesizing information across many documents rather than retrieving a single relevant chunk

A hybrid architecture integrating neural learning with symbolic reasoning, combining pattern recognition with logical deduction and interpretability.

• Symbolic Solvers: Use formal logic, constraint satisfaction, or theorem proving for deductive steps
• Neural Perception: Deep learning handles unstructured inputs like text and images
• GoT Connection: GoT's graph structure mirrors symbolic propositional networks, where nodes represent logical assertions and edges represent inferential relationships
• Interpretability: Unlike pure neural approaches, symbolic components provide auditable reasoning traces suitable for regulated industries

A prompting framework that interleaves discrete reasoning traces with tool-use actions, enabling dynamic strategy updates based on external feedback.

• Thought-Action-Observation Loop: Generates a thought, executes an action (e.g., search), and incorporates the observation into the next reasoning step
• Dynamic Adaptation: Unlike static GoT graphs, ReAct modifies its plan based on real-time environmental feedback
• Complementary Pattern: GoT can serve as the internal reasoning engine within a ReAct agent, structuring the thought generation phase as a graph rather than a linear trace
• Use Case: Interactive tasks requiring API calls, database queries, or web navigation