StePer: Self-Guided Reinforcement Learning for Arithmetic Reasoning
Status: CompletedAuthors: Jiwoo Hong, Huzama Ahmad, Minsu Kim, James Thorne
The project first evaluated how reliably language models could assess their own answers. It then used that signal in a reinforcement-learning loop, rewarding greedy answers that held up against diverse samples from the same model, without a human grader supplying the reward. Across models with one to twenty billion parameters, accuracy improved by up to 5% on four math-reasoning benchmarks. Commonsense reasoning also improved, while general language ability was retained. Human and automated review found the explanations more detailed and logical.