Predict-then-Verify Approaches in LLM Reasoning

Synergy of Training-Based and Training-Free Reasoning Methods

A development team wants to improve a large language model's performance on solving complex logic puzzles without modifying its pre-trained parameters. Their approach involves two stages: first, they prompt the model to generate five distinct potential solutions for a single puzzle. Second, they use an automated checker to evaluate the logical consistency of each of the five generated solutions and select the most valid one as the final output. Which category of training-free reasoning enhancement does this approach primarily represent?

Comparing Training-Free Reasoning Strategies

Match each scenario describing a method to improve a language model's reasoning with the correct training-free approach it exemplifies. Both approaches are applied at inference time without altering the model's pre-trained parameters.