Science Question Answering on ResearchQA (accuracy)
85.8Accuracy (ResearchQA)GPT-OSS-120B
Evaluation Results
| Method | Links | |
|---|---|---|
| GPT-OSS-120BBase Model=GPT-OSS-120B2026.05 | 85.8 | |
| Kimi-K2.5Base Model=Kimi-K2.52026.05 | 84.9 | |
| Deepseek-V3.2Base Model=Deepseek-V3.22026.05 | 81.5 | |
| GLM-5.1Base Model=GLM-5.12026.05 | 80.5 | |
| Qwen3.5-397BBase Model=Qwen3.5-397B2026.05 | 79.4 | |
| OpenAI Deep ResearchBackbone=o3/o4, Sample size=100-sample subset2026.06 | 79.2 | |
| Qwen3.5-122BBase Model=Qwen3.5-122B2026.05 | 79 | |
| Qwen3.5-35BBase Model=Qwen3.5-35B2026.05 | 78.3 | |
| GPT-5 + SearchBackbone=GPT-5, Sample size=100-sample subset2026.06 | 78.2 | |
| Qwen3.5-9B + MaRBase Model=Qwen3.5-9B, Optimization Method=MaR2026.05 | 76 | |
| Perplexity Deep ResearchBackbone=Proprietary2026.06 | 75.3 | |
| DEEPRUBRIC (140-step, ≈750h)Backbone=Qwen3-8B, Steps=140-step, Compute Cost=≈750h2026.06 | 75.2 | |
| Ai2 ScholarQABackbone=Claude Sonnet, Sample size=100-sample subset2026.06 | 75 | |
| Qwen3.5-9BBase Model=Qwen3.5-9B2026.05 | 74.6 | |
| Gemini 3 Pro + SearchBackbone=Gemini 3 Pro2026.06 | 74.3 | |
| DR Tulu-8B (1900-step, ≈9700h)Backbone=Qwen3-8B, Steps=1900-step, Compute Cost=≈9700h2026.06 | 74.3 | |
| WebThinker-32B-DPO (report)Backbone=Qwen2.5-32B (report)2026.06 | 74.2 | |
| DEEPRUBRIC (75-step, ≈402h)Backbone=Qwen3-8B, Steps=75-step, Compute Cost=≈402h2026.06 | 74.2 | |
| Qwen3.5-9B + DAPOBase Model=Qwen3.5-9B, Optimization Method=DAPO2026.05 | 73.4 | |
| WebThinker (report)Backbone=QwQ-32B (report)2026.06 | 72.8 | |
| Qwen3.5-4B + MaRBase Model=Qwen3.5-4B, Optimization Method=MaR2026.05 | 72.7 | |
| Qwen3.5-4BBase Model=Qwen3.5-4B2026.05 | 72.2 | |
| DR Tulu-8B (1000-step, ≈6000h)Backbone=Qwen3-8B, Steps=1000-step, Compute Cost=≈6000h2026.06 | 71.1 | |
| Qwen3.5-4B + DAPOBase Model=Qwen3.5-4B, Optimization Method=DAPO2026.05 | 70.4 | |
| Perplexity-Sonar (High)Backbone=Proprietary2026.06 | 69.1 | |
| Gemini Deep ResearchBackbone=Proprietary2026.06 | 68.5 | |
| DR Tulu-8B (SFT)Backbone=Qwen3-8B2026.06 | 68.5 | |
| Tongyi DeepResearch-30B-A3BBackbone=Qwen2.5-30B2026.06 | 66.7 | |
| WebExplorer-8BBackbone=Qwen2.5-8B2026.06 | 64.8 | |
| DEEPRUBRIC (SFT)Backbone=Qwen3-8B, Stage=SFT2026.06 | 64.8 | |
| Claude-Sonnet SearchBackbone=Claude Sonnet2026.06 | 64.3 | |
| QwQ-32B + RAGBackbone=QwQ-32B2026.06 | 60.9 | |
| Qwen3-8B + RAGBackbone=Qwen3-8B2026.06 | 56.1 | |
| WebThinker-32B-DPOBackbone=Qwen2.5-32B2026.06 | 48.6 | |
| Qwen3-8B + SearchBackbone=Qwen3-8B2026.06 | 46.3 | |
| Search-R1-7BBackbone=Qwen2.5-7B2026.06 | 27.9 | |
| ASearcher-Web-7BBackbone=Qwen2.5-7B2026.06 | 19.4 |