ES-MemEval QA (RAG) - User Modeling: leaderboard

Metric: Token-level F1 (%; answers scored against human-checked reference answers; questions that infer the user's traits, preferences and changing state; 251 questions; bge-m3 retrieval of the four most relevant full sessions from a FAISS index). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.

Top models

#ModelScore
1GPT-3.5 Turbo22
2GPT-4o21.2
3Mistral Small 3.116.4
4Phi-3-medium-128k-instruct13.4
5Ministral-8B-Instruct-241012.6

Interactive version: theaggregate.ai/benchmark?slug=es-memeval-qa-rag-user-modeling · How It Works · Data refreshed daily, snapshot 2026-09-26.