RESEARCH · RESEARCH · #3570
Same Task, Different Statement (arXiv:2611.14997v1): when and how reformulation helps LLM problem solving
This new arXiv preprint tests three types of statement rewrites—Decomposition, Typed Reformulation, and Concept Reduction—on eight math, logic, and code benchmarks with eight models using the same chain-of-thought prompt. Results show large but task-dependent effects: an oracle that picks the best rewrite raises accuracy by +27.6% on average; a simple learned rule transfers part of the gain (+11.0% on problems it chooses to rewrite), while blind rewriting can hurt (−5.6% on a similar share).
KEY POINTS
- This new arXiv preprint tests three types of statement rewrites—Decomposition, Typed Reformulation, and Concept Reduction—on eight math, logic, and code benchmarks with eight models using the same chain-of-thought prompt.
- Results show large but task-dependent effects: an oracle that picks the best rewrite raises accuracy by +27.6% on average; a simple learned rule transfers part of the gain (+11.0% on problems it chooses to rewrite), while blind rewriting can hurt (−5.6% on a similar share).
- The paper shows that targeted statement reformulation can substantially improve LLM reasoning but is not universally helpful, highlighting the need for methods that decide when and how to rewrite.
WHY IT MATTERS
The paper shows that targeted statement reformulation can substantially improve LLM reasoning but is not universally helpful, highlighting the need for methods that decide when and how to rewrite.