Regularization Through Reasoning: Systematic Improvements in Language Model Classification via Explanation-Enhanced Fine-Tuning

Researchers found that fine-tuning large language models with explanations attached to labels significantly improves classification accuracy compared to label-only training. Surprisingly, even random token sequences that mimic explanation structure provide similar benefits, suggesting the improvement comes from increased token budget and regularization rather than semantic meaning.

• →Explanation-enhanced fine-tuning outperformed label-only baselines across 18 dataset and task settings using a 7B-parameter model.
• →Random token sequences mimicking explanation structure provided similar accuracy improvements as genuine explanations.
• →The benefits appear to stem from structural regularization and increased token budget rather than semantic content.
• →Models showed higher activation entropy in intermediate layers and sharper predictive mass at output, indicating more deliberate processing.
• →The findings suggest token-level scaffolding fundamentally shapes how language models perform inference computations.