Tree-of-Thought Beam Search Pruning Lab (3D)
Watch a language model explore a branching tree of reasoning steps: expand candidate thoughts, score them with a value estimate, prune the weak branches, and compare best-first Tree-of-Thought search against a single linear Chain-of-Thought path.
Modern reasoning models don't decode a single chain of tokens and stop — they explore a branching tree of candidate reasoning steps, score each one, and keep only the most promising branches alive, exactly like a best-first search over a state space. This simulator grows that tree in real 3D: pick a problem, expand nodes one step or run the search continuously, and watch a value estimate decide which thoughts survive as the beam narrows toward a solution. Switching to Chain-of-Thought mode collapses the beam to width 1, showing the single committed path a standard (non-reasoning) LLM would generate instead — the same tree, with backtracking removed.
Watch a reasoning model explore a branching tree of candidate thoughts, score each with a value estimate, and prune weak branches with best-first Tree-of-Thought search versus a single linear Chain-of-Thought path.
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install