Rewriting boosts LLMs by 27 percent

🔥 What happened Researchers from eight institutions tested whether LLMs benefit from reformulated problem statements. They evaluated three rewrite types (decomposition, typed reformulation, concept reduction) across eight benchmarks and eight models. 💡 Why it matters An oracle picking the best rewrite after the fact lifts accuracy by 27.6 percent. But blind rewriting loses 5.6 percent — the trick is knowing when to rewrite. ⚡ Our take Prompt engineering just became rewrite roulette: if you don't know when to reformulate, you lose. A simple rule learned from one problem set transfers only +11 percent — that's the real lever.