Reddit users notice long-context 'rot' may be nearly solved within a year

torrid-winnowing · reddit · 2026-09-06

A Reddit poster observes that long-context retrieval in LLMs has improved dramatically in under a year, arguing that 'context rot' — the once-major problem of degraded performance over long contexts — appears to be almost solved, yet the community is barely discussing it. The post is observational without attached benchmarks; readers are encouraged to test current models' long-context performance themselves.

Original post →

More from Models

Models channel →