Evaluating Reasoning in LLMs (2D)

This 2D canvas companion drives the same chain-of-thought scoring model as the 3D version: per-step correctness p = 1 − difficulty, chain probability p^steps, an optional tool-assisted verifier that catches ~55% of chain errors, and a self-consistency majority vote across k samples. A colored grid of benchmark questions resamples every few seconds, and a chain-of-thought walk above it plays out one example reasoning chain step by step.