Why GoT > ToT?
ToT (Yao 2023): Tree β thoughts branch, evaluated, BFS/DFS, but no combining. Each branch isolated.
GoT (Besta 2023/2024): Graph β thoughts can have multiple parents, aggregation nodes combine T1+T2βT4, transformation changes thought, final aggregation. Enables:
β’ Aggregation: T1 (parse) + T2 (calc) β T4 (verified base)
β’ Transformation: T3 (overpay model) β refined
β’ Final aggregation: T4+T5 β T6 (hybrid best)
β’ More expressive than tree, better for complex decisions like mortgage overpay vs invest where you need to combine monthly calc + overpay saving + invest projection.
Byte 4.0: Implements GoT with 6 nodes, 2 aggregations, 1 final, scores 0-1, eye-tracking follows GoT nodes in super face.
Paper: Besta et al. "Graph of Thoughts: Solving Elaborate Problems with Large Language Models" 2023, extended 2024 with aggregation + transformation operations, outperforms ToT on sorting, set intersection, keyword counting, document merging.
Byte 4.0: Same idea for tool workflows β thoughts are tool suggestions, aggregation combines tools.