Brevity Constraints Boost LLM Accuracy

The sharp takeaway is that verbosity is not just style drift, it is a measurable failure mode that can hide latent capability and distort benchmark results.

• –Short-answer constraints appear to recover performance from larger models by reducing overelaboration-induced errors
• –The strongest effect shows up on reasoning-heavy benchmarks like GSM8K and MMLU-STEM, where rank order can reverse
• –For developers, this is a practical prompt lever: shorter outputs can improve accuracy while also cutting token cost and latency
• –The result also weakens naive benchmark comparisons, since response-length policy can change the apparent scaling curve
• –The paper is compelling, but it is still a preprint, so replication across more tasks and deployment settings matters