Reasoning Strategies › Comparing the Strategies
Language AI HandbookAll chapters Reasoning Strategies › Comparing the StrategiesOne section · Sections in this chapter6 / 10 ↓
Reading tools Definitions & inline links
Choose your expertise level to adjust how many terms are explained. Beginners see more tooltips, experts see fewer to maintain reading flow. Hover over underlined terms for instant definitions.
Article links
Make inline references clickable
These four strategies address different failure modes, and the right choice depends on the task. Understanding which strategy to apply requires diagnosing why simpler prompting is failing on your specific problem.
| Strategy | Best for | Core mechanism | Compute cost |
|---|---|---|---|
| Self-consistency | Tasks with verifiable answers | Majority voting over paths | |
| Tree of Thought | Tasks requiring exploration and backtracking | Branching + pruning search | |
| Least-to-most | Long chains with simple sub-problems | Sequential sub-problem solving | sub-problems |
| Decomposed prompting | Mixed tasks needing specialized tools | Tool routing + combination | sub-tasks |
Self-consistency is the right tool when the model produces roughly correct reasoning on most attempts but fails enough of the time to be unreliable. If single-path accuracy is already 90%, self-consistency adds little. If it is 40 to 60%, self-consistency can push it above 80% with 20 to 40 samples. The prerequisite is that answers be normalizable: math answers, letter choices, classification labels, and similar discrete outputs work well. Free-form text answers do not.
Tree of Thought is the right tool when the problem requires exploration: when early decisions are hard to evaluate, when wrong turns must be detected and abandoned, and when the search space is too large for a single path to traverse reliably. The prerequisites are strong: you need a good evaluation function, a well-defined thought decomposition, and enough compute budget for the search. Without these, ToT degrades to expensive beam search with a noisy scorer.
Least-to-most is the right tool when problems are long but composed of simpler pieces. The prerequisite is that decompositions be learnable from few-shot examples and that sub-problems be simpler than the original. For problems where each "sub-problem" is as hard as the original, the method provides no benefit.
Decomposed prompting is the right tool when problems require capabilities the language model lacks: precise arithmetic, factual recall, sorting, or lookup. The prerequisite is that the needed tools be available and that the model can reliably route sub-tasks to the right tools from few-shot examples.
The four strategies are also complementary. Self-consistency can be applied on top of any of the others: run tree of thought or least-to-most multiple times and vote over the final answers. Decomposed prompting can incorporate retrieval as one of its specialized sub-solvers. In practice, state-of-the-art reasoning systems often combine multiple strategies. A production multi-hop question answering system might use DECOMP for task routing, PAL for the arithmetic sub-tasks, and self-consistency over the final answers to reduce variance. Each layer adds reliability at the cost of additional inference.