EntLORE - Cross-Source Composition (Agentic Retrieval): leaderboard
EntLORE (2026) builds an anonymized enterprise world of 2,341 routine documents from an audited truth graph. This board is answer accuracy in percent on its cross-source composition questions (L2), which combine facts stated in several sources, answered through an agent that searches and reads a dense index of the released corpus over at most 30 iterations. Higher is better.
Metric: Answer accuracy (%; L2, 204 questions composing facts stated in several sources; agentic retrieval: search and fetch tools over a dense index of the released corpus, at most 30 LLM iterations; 907 questions over 2,341 anonymized enterprise documents; programmatic scoring for entity, set, count and ordered answers, atomic-claim entailment judged by Claude Opus 4.8 for free-form answers). Source: arxiv.org. Saturation forecast: Around September 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 40.5 | #123 |
| 2 | DeepSeek V4 Pro (Reasoning) | 39.1 | #188 (DeepSeek V4 Pro) |
| 3 | DeepSeek V4 Flash (Reasoning) | 36.4 | #258 (DeepSeek V4 Flash) |
| 4 | GLM-5.2 | 35.9 | #138 |
| 5 | Claude Sonnet 4.6 (Thinking) | 35.6 | #161 (Claude Sonnet 4.6) |
| 6 | Qwen 3.5 397B A17B | 35 | #246 |
| 7 | Kimi K2.6 | 31.5 | #178 |
| 8 | GPT-5.4 Mini | 30.4 | #389 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=entlore-cross-source-composition-agentic-retrieval · How It Works · Data refreshed daily, snapshot 2026-11-13.