graph of thoughts (GoT)

Graph of thoughts generalizes the tree so that partial solutions can not only branch but also merge, loop back, and be refined in place. A tree forces every line of reasoning to stay separate; a graph lets you take two promising partial answers and combine them, or feed one thought through several rounds of self-improvement. Imagine an editor who splits a draft into sections, rewrites each, then stitches the best pieces of several drafts into one.

Formally, thoughts are vertices and directed edges are transformations: generation expands a thought into successors, aggregation merges several thoughts into one, and refinement maps a thought to an improved version (an edge back to itself). A controller schedules these operations and a scoring function ranks vertices. Because aggregation collapses many paths into a single stronger node, GoT can express reasoning topologies that a tree cannot, often reaching comparable quality with fewer total model calls on decomposable tasks.

GoT shines when a task naturally splits into parts whose answers must be combined — sorting, set merging, or document synthesis — where the merge edge does real work. The framework adds engineering overhead: you must define the graph operations, scoring, and schedule, and on tasks that do not decompose the extra machinery buys little over a plain chain or tree.