Results

Average scores per document version, grouped by experiment and retrieval mode. Each score is the percentage of questions answered correctly — higher is better. Not sure what a term means?

Models: claude-haiku = claude-haiku-4-5-20251001 · claude-sonnet = claude-sonnet-4-6 · claude-opus = claude-opus-4-8 · gpt-4o-mini = gpt-4o-mini · gpt-4o = gpt-4o · gpt-5.5 = gpt-5.5 · gemini-2.5-flash = gemini-2.5-flash · gemini-3.5-flash = gemini-3.5-flash