Abstract
CLEAR is a novel approach to language model reasoning that leverages the strengths of a larger (expert) model and smaller (amateur) model. The expert and amateur models each provide feedback on a model's initial output and are contrasted with each other into refined feedback. This feedback is subsequently applied to iteratively improve CLEAR's responses. Similar to how humans would contrast and incorporate multiple feedback they receive to form a high-quality evaluation, CLEAR contrasts expert and amateur model feedback. The method demonstrates notable improvements across multiple domains: up to 19.6% relative increase in story outline interestingness, up to 18.5% increase in constrained generation coverage, up to 6.7% improvement in mathematical reasoning accuracy, and a decrease of up to 22% in toxicity.