Can a Readability Reward Model Fix Incomprehensible AI-Generated Math Proofs?

iScienceLuvr · x · 2026-09-20

Responding to mathematicians' complaints that AI-generated proofs are incomprehensible and don't advance mathematical understanding, the author suggests this looks like a tractable problem for AI labs: build a reward model/judge measuring how easy a proof is to understand, then post-train or steer models with it. The thread debates the catch — readability is reader-dependent, formal rigor and intuitive explanation pull in different directions, and an LLM judge may reward proofs that merely sound clear.

Original post →

More from Models

Models channel →