DecisionRAG Answering - MTRAG Judged: leaderboard
DecisionRAG v1 uses Jev (typesafe/jev-1.13), dense retrieval with Qwen3-Embedding-8B, BGE (bge-reranker-v2-m3), a Qwen3 8B reranker, and GPT 5 with low reasoning. Evaluation populations are MD2D with 4201 validation turns, CQA with 271 evidence-labeled development questions, MTRAG with 188 labeled IBM Cloud turns, and NQ with 500 seeded DAPR test questions. Main MTRAG gold labels are expanded using GPT 5 to judge the pooled top 5 documents; comparisons using the original official labels remain separate. Trained judges use Qwen3 Base at 1.7B, 4B, and 8B; their results are distinct from backbone language-model scores. All source-qualified harnesses and training variants remain separate. Answering uses the top 5 retrieved passages with Qwen3.5-27B FP8, greedy decoding, and thinking disabled. NQ has 334 questions with NQ-open references. MTRAG uses Qwen3.5-27B as its answer judge. Oracle Gold Passages is an oracle control, not an ordinary system. Metric: Accuracy (%).
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DecisionRAG Qwen3.5-27B FP8 no-thinking answerer + dense candidates | 47.3 | |
| 2 | DecisionRAG Qwen3.5-27B FP8 no-thinking answerer DecisionRAG (Jev) | 44.1 | |
| 3 | DecisionRAG Qwen3.5-27B FP8 no-thinking answerer Dense top-100 + Qwen3-RR | 39.4 | |
| 4 | DecisionRAG Qwen3.5-27B FP8 no-thinking answerer Dense top-20 + GPT-5 | 34 | |
| 5 | DecisionRAG Qwen3.5-27B FP8 no-thinking answerer Dense | 31.4 | |
| 6 | DecisionRAG Qwen3.5-27B FP8 no-thinking answerer None | 6.4 |
Interactive version: theaggregate.ai/benchmark?slug=decisionrag-answering-mtrag-judged · How It Works · Data refreshed daily, snapshot 2026-10-28.