Larger and more instructable language models become less reliable

Large language models (LLMs) have been scaled up and shaped up (via instruction tuning, RLHF, moderation) to increase capability and instructability. Studying GPT, LLaMA and BLOOM families across five benchmarks and human studies, the authors find that while scaled and shaped models improve correctness and prompt stability, they become less reliable for users: errors persist on instances that humans consider easy (difficulty discordance), avoidance is reduced and replaced by seemingly plausible but wrong answers (ultracrepidarianism), and stability improvements still leave pockets of variability across difficulty levels. Human supervisors often fail to detect these errors, especially on harder items. The work argues for rethinking design toward difficulty-aware reliability and calibrated avoidance, particularly for high-stakes applications.

Lexin Zhou, Wout Schellaert, Fernando Martínez-Plumed, Yael Moros-Daval, Cèsar Ferri, José Hernández-Orallo