Smaller language models outperform giant ones on 7.7% of tasks when forced to answer briefly

Researchers found that the largest language models (like GPT-4) actually perform worse than much smaller models on roughly 1 in 13 benchmark problems, because they tend to over-explain their answers in ways that introduce errors. Adding a simple instruction to keep responses short reverses these performance gaps entirely, with the largest models now outperforming small ones by 7-16 percentage points on math and science questions.

This suggests that AI evaluation practices are systematically masking the actual capabilities of larger models — the bigger models aren't fundamentally weaker on certain tasks, they're just being asked to respond in ways that trip them up. The practical implication is immediate: how you ask an AI to answer a question matters as much as which model you use, and companies deploying large language models may be wasting computational resources by using one-size-fits-all prompts instead of tailoring how they phrase requests. This inverts the common assumption that you simply need a bigger model to get better results.