ES-MemEval QA (Long Context) - Temporal Reasoning: leaderboard
Questions on the order of events and how they depend on each other over time (236 questions). ES-MemEval (Chen et al., Tongji University, 2026) asks questions about the long emotional-support chat histories of 18 virtual users (about 22 sessions and 13K tokens each), scored by token-level F1 in percent against human-checked reference answers; higher is better. The model answers from the dialogue history within the context window the paper gave it: 128K tokens for the open models, 4K for gpt-3.5-turbo and 16K for gpt-4o.
Metric: Token-level F1 (%; answers scored against human-checked reference answers; questions that order events and follow how they depend on each other over time; 236 questions; no retrieval: the dialogue history (11K-19K tokens per user) within the context window the paper gave each model, 128K tokens for the open models, 4K for gpt-3.5-turbo and 16K for gpt-4o). Source: arxiv.org. Saturation forecast: Estimated already saturated. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-3.5 Turbo | 22.4 | #2773 |
| 2 | GPT-4o | 19.6 | #741 |
| 3 | Mistral Small 3.1 | 18 | #1699 |
| 4 | Phi-3-medium-128k-instruct | 12.8 | #3623 |
| 5 | Ministral-8B-Instruct-2410 | 9.3 | #4720 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=es-memeval-qa-long-context-temporal-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-12.