[jerjerjer]

Is there a benchmark to measure real effective context length?

Sure, gpt-4o has a context window of 128k, but it loses a lot from the beginning/middle.

[brookst]

Here's an older study that includes Claude 3.5: https://www.databricks.com/blog/long-context-rag-capabilitie...?

[evertedsphere]

ruler https://arxiv.org/abs/2404.06654

nolima https://arxiv.org/abs/2502.05167

[bigmadshoe]

They often publish "needle in a haystack" benchmarks that look very good, but my subjective experience with a large context is always bad. Maybe we need better benchmarks.