Question Answering on HotpotQA (test) (EM Metric Only)
42.5EMSD-Search-Base
Evaluation Results
| Method | Links | |
|---|---|---|
| SD-Search-BaseTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 42.5 | |
| SD-Search-InstructTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 42.4 | |
| MR-Search-BaseTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 41.9 | |
| AutoRefine-BaseTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 40.5 | |
| AutoRefine-InstructTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 40.4 | |
| Thinker-InstructTraining Pipeline=Synthesize → SFT, External Teacher=Qwen2.5-72B, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 40 | |
| ReSearch-InstructTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 35.1 | |
| StepSearch-InstructTraining Pipeline=Pure RL, External Teacher=GPT-4o, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 34.5 | |
| Search-R1-InstructTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 33.1 | |
| StepSearch-BaseTraining Pipeline=Pure RL, External Teacher=GPT-4o, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 32.9 | |
| ReSearch-BaseTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 30.5 | |
| Search-R1-BaseTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 29.7 | |
| Naive RAGTraining Pipeline=Zero-Shot, External Teacher=–, Retrieval Strategy=w/ Single-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 25.5 | |
| Search-o1Training Pipeline=Prompting, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 22.1 | |
| R1-InstructTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/o Retrieval, Backbone=Qwen2.5-3B2026.05 | 20.8 | |
| R1-BaseTraining Pipeline=Pure RL, External Teacher=–, Retrieval Strategy=w/o Retrieval, Backbone=Qwen2.5-3B2026.05 | 20.1 | |
| SFTTraining Pipeline=SFT, External Teacher=–, Retrieval Strategy=w/o Retrieval, Backbone=Qwen2.5-3B2026.05 | 18.6 | |
| IRCoTTraining Pipeline=Prompting, External Teacher=–, Retrieval Strategy=w/ Multi-Hop Retrieval, Backbone=Qwen2.5-3B2026.05 | 16.4 | |
| Direct GenerationTraining Pipeline=Zero-Shot, External Teacher=–, Retrieval Strategy=w/o Retrieval, Backbone=Qwen2.5-3B2026.05 | 14.9 |