AI predicts crowd highlights from text alone, beating lead baseline by 5.5%

A new paper from researchers Kazuki Nakayashiki and Keisuke Watanabe tackles the cold-start prediction of crowd highlight salience: can a model predict which passages readers will mark before any marks have been made? Prior work found that zero-shot language models performed worse than a simple 'lead baseline' (assuming the first sentences are most salient). The authors trained a logistic ranker on sentence embeddings combined with positional and contextual features, using a pre-registered ladder of models and a by-document cluster bootstrap. The trained model beats the lead baseline by +0.044 average precision (95% CI [+0.029, +0.058]), clearing a pre-registered margin delta=0.03 in 97% of resamples. In product terms, precision@3 jumps from 0.25 to 0.39 (+55% relative), and the model outperforms the lead baseline on 69% of documents.

An ablation study attributes the edge to raw embeddings (+0.014) and training augmentation (+0.010). The advantage is not explained by temporal generalization failure, content drift, or near-duplicate leakage. Standardized regression reveals the edge is governed mainly by document popularity: the model shines on less popular 'long tail' content, where the lead baseline weakens, while on extremely popular content the lead baseline itself becomes stronger. Two unsupervised extractive baselines (centroid and LexRank-style centrality) lose to the lead baseline, confirming the trained model genuinely learns from real reader marks rather than generic text properties. The work demonstrates that a relatively simple trained model can predict crowd salience reliably for new documents, with practical implications for highlighting tools, summarization, and content recommendation systems.