Reformulating Problem Statements Boosts LLM Accuracy, Study Finds

When researchers rewrote problem statements instead of changing the model or its reasoning chain, they observed large accuracy jumps for large language models (LLMs). The finding matters because it highlights a low‑cost lever—statement reformulation—that creators and developers can apply without retraining or adding external tools.

What the study set out to test

The Grafy Research Lab paper, "Same Task, Different Statement: When and How Reformulation Helps LLM Problem Solving" (arXiv, 2026), asks whether the human habit of restating a hard problem can improve LLM performance. The authors treat a rewrite as legal only if it preserves the original meaning, constraints, goal, answer interface, and the set of correct solutions (Section 3). They then evaluate three families of rewrites on a suite of benchmarks.

How reformulation was formalized

A problem is represented as a pair $(x, g)$ where $x$ is the statement and $g$ the solving goal. A reformulation produces a new pair $(x', g')$ that satisfies the task‑preservation constraints (Section 3.1). The transformation is isolated from the solver: all models receive the same zero‑shot chain‑of‑thought (CoT) prompt, and the only variable is the statement variant. This design lets the authors attribute any performance change purely to the representation.

The three rewrite families

Family	What it does	Typical target problem type
Decomposition (D)	Makes hidden subgoals explicit, often by breaking a story into separate premises.	Search, Classification
Typed Reformulation (TR)	Casts the task into a more suitable solving frame (e.g., turning a design question into an optimization formulation).	Optimization, Design, Verification
Concept Reduction (C)	Strips surface details to expose the underlying logical or mathematical structure.	Explanation, Causality

Each family respects the legality constraints, and the authors generated reformulations automatically then audited them for correctness.

How much accuracy improved

The experiments covered eight benchmarks (math, logic, code) with a total of 1,064 analysis problems and a held‑out set of the same size. Eight LLMs—ranging from open‑source to proprietary models—were tested under the identical CoT prompt.

• Baseline accuracy on original statements varied between 43.7% and 63.6% across models.
• An oracle that retrospectively picks the best legal rewrite for each problem lifts accuracy to 80.0%–86.4%, a gain of $+27.6\%$ on average.
• A simple rule learned from one problem set, then applied to unseen problems, achieved a $+11.0\%$ lift on the subset it chose to rewrite, while blind rewriting (rewriting every problem regardless of type) caused a $5.6\%$ drop.

These numbers show that the right rewrite can close a large portion of the gap between current LLM performance and the theoretical optimum for a given task.

When each rewrite helps—or hurts

The authors broke down performance by problem type (nine categories identified by LLMs and validated by humans). The patterns were clear:

• Decomposition helped most with Search and Classification tasks, where exposing sub‑questions guides the reasoning flow.
• Concept Reduction shone on Explanation and Causality problems, where removing narrative fluff reduces cognitive load.
• Typed Reformulation was advantageous for Optimization, Design, and Verification problems, because recasting the goal aligns with the model’s internal problem‑solving heuristics.
• A mismatch—applying the wrong family—could actually lower accuracy, confirming that reformulation is not a universal fix.

Practical implications for developers and researchers

The study suggests a workflow that many Grafy users can adopt today:

1. Identify the problem type – Use the same heuristics the authors employed (e.g., look for keywords like "search", "optimize", or "explain").
2. Pick a rewrite family – Apply the corresponding transformation: break the statement into steps, change the framing, or strip extraneous details.
3. Run the LLM with the standard CoT prompt – No changes to model parameters or additional tools are required.
4. Evaluate – If the answer is wrong, try a different family; the paper shows that a single well‑chosen rewrite often flips the result.

Grafy users can prototype this loop with Grafy Chat to experiment with different prompt variants, and Grafy Reader to load the full paper, explore the released dataset, and even query the authors’ audit logs. For more systematic tracking, the Grafy Graph app lets you build a branching tree of original statements, each rewrite, and the model’s outputs, making it easy to compare which transformation worked best for a given benchmark.

Limitations and next steps

The authors note three key caveats:

• The study only examined zero‑shot CoT prompting; other solving strategies (self‑refine, tool use) were not combined with reformulation.
• The datasets, while diverse, still reflect a research‑grade selection; real‑world user queries may exhibit different distributions.
• The simple selection rule was trained on one set of problems; more sophisticated meta‑learners could further improve transferability.

Future work could integrate reformulation directly into the prompting pipeline—e.g., a pre‑processor that automatically classifies a query and rewrites it—potentially turning the observed gains into a seamless feature for LLM‑powered applications.

---

Source: "Same Task, Different Statement: When and How Reformulation Helps LLM Problem Solving," arXiv, 2026‑11‑14. https://arxiv.org/abs/2611.14997

Frequently asked questions

Does the study claim that any rewrite always improves LLM results?

No. The authors found that a mismatched rewrite can hurt performance, and that only certain families help specific problem types. The overall gain comes from selecting the appropriate rewrite, not from applying rewrites indiscriminately.

Can I use Grafy to automatically apply these reformulations?

Grafy does not yet provide an out‑of‑the‑box automatic reformulation engine, but you can manually experiment with the three families using Grafy Chat and track results with Grafy Graph. The released dataset and scripts are also accessible via Grafy Reader.

How large was the accuracy improvement when the best rewrite was chosen?

When an oracle picked the best legal rewrite for each problem, accuracy rose by $+27.6\%$ on average, moving from a baseline of roughly 44‑64% up to 80‑86% across the eight models tested.

---

Source: Grafy Research Lab — Same Task, Different Statement: When and How Reformulation Helps LLM Problem Solving