Large language models can make technical content more accessible, but their outputs have a hard ceiling when it comes to interactive learning. Because LLM explanations are static text, readers cannot execute a step, test an assumption, or see how a proof unfolds dynamically. A new arXiv paper, "Explorable Proofs," identifies this limitation and argues for a different approach.

The authors suggest grounding written proofs in formal representations—machine-readable structures that capture the logical skeleton of an argument. By linking prose to these formal objects, a proof could become explorable: readers could step through it, inspect each inference, and even manipulate parts of the proof to see what changes. This would turn a passive reading experience into an active investigation.

The paper does not present a full implementation, but it makes a conceptual case for why such grounding is necessary. If successful, this direction could help students and researchers move beyond trusting an explanation to actually verifying and understanding the underlying mathematics.