arXivAryo Pradipta Gema, Beatrice Alex, Pasquale MinerviniWed, Jul 1, 2026, 7:41 AM PDT
score 17.0
New method finds AI heads that retrieve non-literal meanings
Original: Logit-Contribution Scoring Identifies Non-Literal Retrieval Heads
Source: arxiv.org ↗
Writing ELI5 summary…