SourceNumber Cookbook: Number Understanding of Language Models and How to Improve It
The paper observes that within the same model family, the larger and smaller variants achieve nearly identical NUPA scores across most tasks. GPT-4o and GPT-4o-mini perform almost identically, and Qwen2-72B and Qwen2-7B show a similar pattern. The authors conclude that once a model reaches a certain size, NUPA performance depends more on architecture, training strategies, data diversity, and post-training refinements rather than on simply increasing parameter count. This contrasts with the general trend in LLMs where larger models outperform smaller ones on most benchmarks.