Brevity Constraints Reverse Performance Hierarchies in Language Models

We evaluate 31 language models (0.5B–405B parameters) across 1,485 problemsfrom five standard benchmarks and identify a systematic but correctable failuremode: on 7.7% of problems, small models (≤10B) outperform large models (≥70B)by an average of 28.4 percentage points (Cohen's d = 1.34).

The mechanism is scale-dependent verbosity. Large models spontaneouslygenerate responses 59% longer than small models on these problems — not throughmore explicit reasoning steps (9.1 vs 10.5), but through verbose implicitelaboration that accumulates errors. We call this overthinking.

A simple intervention reverses the hierarchy. Adding brevity constraints("answer in under 50 words") improves large model accuracy by +26.3pp andreduces the performance gap by 67%. Critically, on GSM8K and MMLU-STEM, thegap doesn't just close — it fully reverses: large models go from losing by13.1pp and 27.3pp to winning by 7.7pp and 15.9pp respectively.

This effect is architecture-independent, replicating across Llama, Qwen,Gemma, and Mistral families (5/5 datasets each), and operates continuouslyacross the full parameter spectrum (Pearson r = −0.388, p = 0.0035).

Three independent contamination tests (response diversity: 89–100% unique;length variability: CV = 0.31–1.21; error taxonomy: 41–82% over-reasoningfailures) confirm genuine capability differences rather than memorizationartifacts.

The implication: inverse scaling on standard benchmarks reflects promptdesign failure, not architectural limitation. Large models possess superiorlatent capabilities that universal prompting obscures. Scale-aware promptengineering — not larger models or retraining — is sufficient to recover them.