LifeAgentBench - Database-Augmented Prompting: leaderboard
LifeAgentBench (Tian et al., EMNLP 2026) asks questions over long-horizon diet, sleep, physical-activity and emotion records of 100 AI4FoodDB participants, with answers computed by executable queries and scored by exact match (numeric answers within a small tolerance). In database-augmented prompting the model writes a read-only SQL query against the records database, the system executes it and the model answers from the returned evidence, over single- and multi-user questions. Accuracy in percent; higher is better.
Metric: Answer Accuracy (%; all questions, model-written read-only SQL retrieval then answer). Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro | 55.67 | #181 |
| 2 | Gemini 2.5 Flash Lite | 39.04 | #958 |
| 3 | GPT-4o | 34.71 | #741 |
| 4 | Claude 3 Haiku | 29.3 | #2629 |
| 5 | Qwen 3.5 9B | 25.76 | #803 |
| 6 | Llama 3.1 8B Instruct | 21.53 | #4277 |
| 7 | Phi-3.5-mini-instruct | 16.16 | #5757 |
| 8 | Gemma 2 9B (IT) | 14.54 | #2275 |
| 9 | Llama 3.2 3B Instruct | 13.47 | #7185 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=lifeagentbench-database-augmented-prompting · How It Works · Data refreshed daily, snapshot 2026-11-12.