Tree of Thoughts Prompting: How to Get 4%→74% Reasoning Gains With Plain Prompts
bigaiguy · x · 2026-10-09
Princeton and Google DeepMind researchers developed Tree of Thoughts (ToT), a prompting technique where the model explores multiple reasoning branches, scores them, and backtracks — lifting accuracy on hard reasoning tasks from 4% to 74% in the original paper.
The four-step framework: thought decomposition, thought generation (multiple options per step), state evaluation (sure/maybe/impossible), and search with backtracking.
The thread includes five copy-paste ToT prompt templates: writing (5 scored opening angles, prune below 7, expand top 2), strategy (three opposing strategists propose 90-day plans, critique each other, merge the best), debugging (generate 4 ranked root causes with exact confirmation tests), decision-making (3 approaches with best/worst cases and success probability), and the classic three-experts universal template. No special model needed — just the right prompt.
More from Apps
- ChatGPT Finances: synced bank data isn't trained on, but your chats about it are — Ok_Negotiation_2587 · 2026-10-09
- Dev ships Raycast extension for RevenueCat with official MCP and AI queries — samuelkraft · 2026-10-09
- Skip annual AI subscriptions: pay monthly first, plus free-tool picks like Photopea — sujingshen · 2026-10-09
- AI assistants should tackle cost of living, not vacations, argues product veteran Jim Prosser — sebpaquet · 2026-10-09
- Google AI Pro/Ultra Subscribers Get Monthly Gemini API Credits — Don't Forget to Redeem — danielrock · 2026-10-09
- She builds household utility apps and markets them to her partner with Claude-made promo videos — misraturp · 2026-10-09