Revisiting Generalization Across Difficulty Levels: It's Not So Easy

Yeganeh Kordi, Nihal V. Nayak, Max Zuo +2

We investigate how well large language models (LLMs) generalize across different task difficulties, a key question for effective data curation and evaluation. Existing research is…