Two ways to trade extra compute for more reliable answers — sample several reasoning paths and take the majority vote, or have the model branch, evaluate, and prune ideas like a search tree.
Why: one chain of thought can take a wrong turn; running several and taking the most common answer cancels out the flukes. When: use it for high-stakes reasoning where being right matters more than cost. How: raise temperature so paths differ, run 5 times, count the answers.
Set temperature to about 0.7 and run this prompt 5 separate times:
"Three friends split a $90 bill, but one only had a salad and
pays half what each of the others pays. How much does each pay?
Think step by step, then give the final amounts."
Then take the answer that appears most often as your result.Why: some problems need exploring several options before committing — Tree of Thoughts has the model propose multiple next steps, score them, and continue only from the best. When: use it for planning, puzzles, or design problems with many viable paths.
Solve this as a tree of thoughts.
Goal: plan a 3-stop day trip in Rome on foot.
Step 1: propose 3 different opening stops.
Step 2: for each, briefly rate walking distance and crowd level (1-5).
Step 3: keep the best opening, then propose 3 possible second stops.
Step 4: pick the best full route and explain why you pruned the others.Why: both techniques call the model many times, so they cost several times more and run slower. When: reserve them for the few prompts where reliability is worth the price; for everyday tasks, a single chain of thought is plenty. Where: note the token cost in your prompt log so the tradeoff stays visible.
Rule of thumb:
Single answer -> cheap, fast, fine for most tasks
Chain of Thought -> one call, big accuracy gain on reasoning
Self-consistency -> ~5x cost, use for high-stakes answers
Tree of Thoughts -> highest cost, use for search/planning
Always start cheap and only escalate if the answer is unreliable.