OAKS-BABI (RAG): leaderboard
Metric: Interval-level accuracy (%) on OAKS-BABI (1,224 questions over bAbI-style synthetic stories whose facts change across 2K-token chunks), all questions: the same questions are asked after every chunk and each answer is scored against the answer valid at that point, averaged over intervals and questions; RAG setting (top 30 earlier chunks retrieved with Qwen3-Embedding-0.6B); higher is better. Source: arxiv.org. Saturation forecast: Around March 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 62.2 | #145 |
| 2 | Gemini 2.5 Flash | 56.6 | #237 |
| 3 | Gemini 3 Pro | 51.3 | #77 |
| 4 | Qwen 3 235B A22B 2507 Instruct | 44.7 | #291 |
| 5 | Qwen 3 Next 80B A3B Instruct | 43.5 | #352 |
| 6 | Gemma 3 27B (IT) | 38.6 | #509 |
| 7 | Qwen 3 30B A3B 2507 Instruct | 37.8 | #464 |
| 8 | Qwen 3 8B (Non-reasoning) | 35.1 | #667 (Qwen 3 8B) |
| 9 | GPT-OSS-120B | 33.7 | #330 |
| 10 | Qwen 2.5 7B Instruct | 32.4 | #846 |
| 11 | Qwen 3 4B 2507 Instruct | 29.3 | #745 |
| 12 | Gemma 3 4B (IT) | 25.2 | #971 |
| 13 | GPT-OSS-20B | 22.2 | #499 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=oaks-babi-rag · How It Works · Data refreshed daily, snapshot 2026-10-11.