Beyond Benchmark Scores: How Data Distribution Shapes True LLM Capability

This research reveals a critical disconnect between benchmark leaderboard dominance and real-world LLM performance. By isolating data distribution as the primary variable, the study demonstrates that models optimized for benchmark-aligned data develop brittle, narrow representations that fail on out-of-distribution tasks. The findings challenge industry incentives that prioritize benchmark scores and offer diagnostic tools to identify 'benchmark shadows' in model parameter spaces.