Same Task, Different Statement: When and How Reformulation Helps LLM Problem Solving
Summary
- Researchers investigate whether rewriting problem statements helps large language models solve complex tasks across eight math, logic, and code benchmarks.
- Three rewrite strategies are evaluated: Decomposition (making subgoals explicit), Typed Reformulation (casting into a better solving frame), and Concept Reduction (removing surface detail).
- An optimal oracle rewrite yields an average +27.6% accuracy increase, though mismatched rewrites can hurt performance.
- A simple fixed rule learned from one set of problems can selectively transfer gains to new problems, improving accuracy by +11.0% on selected problems.
Why it matters
Understanding when and how to reformulate problem statements enables more reliable LLM problem solving, moving beyond blind rewriting to targeted, context-aware prompt engineering.