Reasoning-Grounded Natural Language Explanations for Language Models

Reasoning processes can make AI explanations more trustworthy: The paper demonstrates that when language models ground their explanations in an explicit reasoning process, the explanations become more faithful to what the model actually did, rather than post-hoc justifications that may not reflect the true decision-making.

Separating reasoning from answers improves honesty: By using a "joint predict-explain" approach where answers and explanations are derived independently from the same reasoning sequence (rather than explanations being generated after answers), the model reduces the risk of explanations being tailored to justify predetermined answers.

Better explanations can also mean better answers: The research shows that incorporating reasoning processes doesn't just improve explainability—it also enhances the quality of the model's actual answers across multiple problem domains, suggesting these goals are complementary rather than trade-offs.

Language models often reuse their reasoning steps: The paper observes that models frequently copy elements directly from their reasoning sequences into both final answers and explanations, indicating that transparent reasoning processes provide genuine insight into model behavior rather than obscuring it.

This approach works across different domains: The technique demonstrates "high alignment between answers and explanations in several problem domains," suggesting the method is broadly applicable rather than limited to specific use cases.