LakeQA: leaderboard
LakeQA (2026) tests LLM agents on exploratory multi-hop question answering over a data lake of about 40 million Data.gov and Wikipedia documents, where the agent must search, download and query the relevant files before answering. The board is exact match in percent on the 1,007 LakeQA-full tasks; higher is better.
Metric: End-to-end exact match (%) on the 1,007 LakeQA-full tasks; the agent answers a multi-hop question over a data lake of about 40 million Data.gov and Wikipedia documents with one tool call per round (search, list, download, inspect, query) until it answers or hits the turn limit; exact match against the annotator-verified answer; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 32.87 | #250 |
| 2 | GPT-5.2 | 18.37 | #176 |
| 3 | DeepSeek R1 | 15.99 | #473 |
| 4 | Claude Haiku 4.5 | 11.42 | #568 |
| 5 | GPT-5 Mini | 5.16 | #348 |
| 6 | Llama 3.3 70B Instruct | 5.06 | #1357 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=lakeqa · How It Works · Data refreshed daily, snapshot 2026-11-13.