EMNLP paper with Sebastian Padó: We show that surprisal-based transformer predictions for human reading times diverge for some agreement attraction configurations (ORC) but not others (PP-mod). We discuss theoretical and methodological implications. arxiv.org/abs/2603.16574