Brevity Constraints Reverse Performance Hierarchies in Language Models

This preprint investigates how imposing brevity constraints on language model outputs can unexpectedly reverse established performance hierarchies across model families and sizes. The findings have significant implications for understanding test-time compute efficiency and the deployment behaviour of large language models under constrained generation settings.

Author:

• MD Azizul Hakim (Lead Author, Bangladesh Sweden Polytechnic Institute, Bangladesh)

arXiv: 2604.00025

Key Contributions:

• Demonstrated that brevity constraints can reverse conventional performance rankings among LLMs
• Provided empirical evidence challenging assumptions about model size and output quality
• Advanced understanding of inference-time behaviour and test-time compute efficiency in LLMs
• Sole-author publication demonstrating independent research capability

Recommended citation: Hakim, MD Azizul (2025). "Brevity Constraints Reverse Performance Hierarchies in Language Models." arXiv preprint. arXiv:2604.00025
Download Paper