'Emotional Support' Prompts Significantly Boost LLM Math Reasoning

gajesh · x · 2026-08-11

A developer analyzing Claude's progress on the Riemann hypothesis noted an intriguing finding: most of the effective inputs weren't math-related, but rather positive encouragements telling the model to "keep going."

Reflecting on their own experience optimizing MLX, the developer uses a loop to regularly deliver a fixed script of encouragement, claiming that major breakthroughs are possible. They found that over 90% of improvements occur either during or immediately after the encouragement round, showing astonishing parallels to human cognitive behavior.

Original post →

More from Fun

Fun channel →