Reasoning Strategies › Summary

This chapter covered four reasoning strategies that improve language model performance beyond simple prompting:

• Self-consistency samples $N$ diverse reasoning paths and votes on the most common answer. It exploits the asymmetry between correct answers (which concentrate) and wrong answers (which scatter). It is effective on verifiable tasks and achieves large gains with 20 to 40 samples, at the cost of $N\times$ inference compute.
• Tree of Thought treats reasoning as a search problem, generating and evaluating multiple candidate thoughts at each step, then pruning unpromising branches. It enables explicit exploration and backtracking, dramatically improving performance on structured tasks like puzzles where early decisions determine later possibilities.
• Least-to-most prompting decomposes problems into a sequence of sub-problems from simplest to most complex, solving them in order and carrying results forward. It separates problem analysis (stage 1) from execution (stage 2), and is particularly powerful for tasks with long dependency chains and for overcoming length generalization failures.
• Decomposed prompting strategies generalize decomposition by routing sub-tasks to specialized solvers: tools for computation, retrievers for facts, interpreters for programs. They allow the language model to focus on reasoning and coordination while delegating computation and factual recall to systems that handle them exactly.

These strategies are complementary and often combined in practice. The underlying insight they share is that language models reason more reliably when the task is structured to match their strengths: generating coherent intermediate steps rather than leaping to conclusions, exploring alternatives rather than committing to a single path, and solving familiar sub-problems rather than novel complex ones at the limit of their reasoning capacity.

The broader lesson is that inference-time compute can substitute for model capability. A weaker model using self-consistency, ToT, or decomposition can match a stronger model using direct prompting, by spending more compute at inference time. This trade-off between model size and inference strategy is a recurring theme in modern language model deployment, and understanding it is essential for building systems that are both capable and efficient.