Evaluating Fine-Tuned LLMs on Reasoning Puzzles

This article evaluates how fine-tuning affects AI reasoning on structured puzzle tasks. Using Open-LLaMA as a base, models were trained on datasets of varying sizes (1M, 10M, 100M). Results show clear scaling benefits: the 100M-sample model achieved the best pass@1 accuracy in both in-distribution and out-of-distribution tests. While smaller models struggled with limited reasoning steps or logical errors, larger fine-tuned models demonstrated deeper problem-solving ability, outperforming both base and prompt-engineered approaches.