Search over thoughts
Instead of one chain, grow a tree of candidate thoughts, score them, and search. Tree of Thoughts, best-of-N, and MCTS-style deliberation — and their cost.
A single chain-of-thought commits to one path and walks it to the end, even when an early step was a dead end. Search over thoughts instead grows several candidate next-steps, scores them, expands the promising ones, and can backtrack — turning deliberation into the same explore-and-exploit problem that made game-playing systems strong.
Start here
One chain is a single sample. A tree is a search: generate several candidate thoughts, ask which look promising, keep the good branches, and drop the rest. Backtracking is the payoff — the ability to abandon a wrong turn instead of reasoning further down it.
One hard problem, one model — only the thinking budget changes
Problem: "A farmer has 24 apples in three baskets. The second basket has twice the first; the third equals the first. How many in each?"
sets up equations
The characteristic shape: steep early gains as the budget buys actual setup and self-checking, then a plateau — extra thinking rarely rescues a model past its ceiling. Curve and samples are illustrative; on easy problems the curve is flat from token zero, and the cost is paid in latency whether or not accuracy moves.
The underlying lever is the same: more test-time compute, spent better. Search is what "spend it better" looks like when a single chain plateaus.
Tree of Thoughts
- Generate — from a state, produce several candidate next thoughts.
- Evaluate — score each candidate, either by the model or a value function. This step is the weak link: a bad evaluator searches badly.
- Select and expand — keep the best, discard the rest, recurse.
- Backtrack — if a branch stalls, return to an earlier node and try another. This is what a single chain cannot do.
Search strategies
- Best-of-N — sample N complete answers, pick the best. Parallel, simple, no backtracking.
- BFS / DFS over thoughts — breadth or depth-first exploration of the candidate tree, guided by the evaluator.
- MCTS-style (LATS) — simulate, score, and back-propagate values to balance exploring new branches against exploiting good ones, unifying reasoning, acting, and planning in one tree.
Careful
Search multiplies cost — N branches means roughly N× the tokens — and it is only as good as its evaluator. If the model cannot tell a good partial thought from a bad one, a bigger tree just explores more of the wrong space. On many tasks a single long chain beats an elaborate search; reach for the tree when the task genuinely branches and the branches are scoreable.
Check yourself
Eduspheria wiki · Agentic AI, Planning & reasoning
0 / 5 answered
Next: what to do when the plan itself is wrong.