FoT：動態LLM推理優化框架

🔑 增強重點摘要

• •FoT builds on established reasoning paradigms like Chain-of-Thought (CoT) (linear structure, introduced 2022), Tree-of-Thoughts (ToT) (branching exploration, 2023), and Graph-of-Thoughts (GoT) (general graphs with cycles and merging), addressing their static limitations through dynamic schemes.[1][4][5]
• •Preceding works highlight ToT's tree-like fan-shaped embeddings and GoT's superior expressiveness for multi-path interactions and cyclicity, which FoT optimizes with hyperparameter tuning and caching for faster execution.[1]
• •FoT implements and improves ToT, GoT, and ProbTree, achieving reduced costs and better benchmarks via parallel execution and prompt optimization, extending graph-structured reasoning trends.[1][2]
• •Historical shift from linear CoT to DAGs and graphs recognizes real reasoning's branching, merging, and reuse, which linear prompts obscure; FoT enables adaptable frameworks beyond fixed topologies.[2][4]
• •Open-source FoT codebase supports future developments, aligning with inference-time scaling techniques like search over paths and self-refinement that boost LLM reasoning without retraining.[5]

🛠️ 技術深入

• •FoT overcomes static topologies: CoT (linear, simplest, lowest accuracy), ToT (tree-like radial embeddings in high-dim space, branching without merging), GoT (graph Laplacian encoding, supports cycles/multi-paths, higher H1 persistence for mesh-like organization).[1]
• •Graph structures enable premise reuse, branching/merging absent in chains; DAGs make dependencies explicit vs CoT's arbitrary linearization.[2]
• •Inference-time scaling context: FoT aligns with search over paths (ToT/GoT style), using more tokens/time for exploration correlating with correctness, no model retraining needed.[1][5]
• •ToT uses model-intrinsic evaluation (vs GoT's self-scoring); FoT adds hyperparams/prompt opt for efficiency.[4][5]

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線