Modarressi et al. present NoLiMa
Presented NoLiMa — in a long-context evaluation with literal overlap removed, many models dropped below 50% of their short-context performance at 32K
Why it matters
Showed that needle-in-a-haystack scores overstate long-context ability
Source — Primary sources first; secondary reports are labelled “Press”
Paperarxiv.orghttps://arxiv.org/abs/2502.05167Added in Week 35, 2026 · See that week