Reasoning sketches, at a glance.

Humans naturally compress complex reasoning into crisp sketches. Sketch-of-Thought (SoT) channels that instinct for language models, guiding them to draft conceptual scaffolds before final answers. Explore the cached prompts below to see how sketching trades fewer tokens for faster, still dependable reasoning.

GitHub Repo Read the Paper

Task TypeChoose a reasoning domain

Reasoning ModesSoT paradigms vs. CoT baseline

Pick a reasoning mode to see how the sketch operates.

Token Usage—

Current output— tokens

Select a paradigm to compare against the Chain-of-Thought baseline.

Latency—

Measured latency— ms

Estimated rate at a decoding speed of 35 tkn/s.

—

Select a task to explore.

Reasoning Trace

No <think> section captured yet.

Final Answer

—