Multi-hop Question Answering on HotPotQA (Acc, TTFT, Delay (s))
60.3AccuracyStreamingThinker
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| StreamingThinkerParadigm=Streaming, Reasoning Depth=D3, Model=Qwen3-4B2025.10 | 60.3 | 24.32 | 6.5 | |
| InterleavedParadigm=Interleaved, Reasoning Depth=D3, Model=Qwen3-4B2025.10 | 59.8 | 24.32 | 14.45 | |
| StreamingThinkerParadigm=Streaming, Reasoning Depth=D2, Model=Qwen3-4B2025.10 | 58.1 | 24.32 | 3.92 | |
| InterleavedParadigm=Interleaved, Reasoning Depth=D2, Model=Qwen3-4B2025.10 | 57.6 | 24.32 | 12.02 | |
| BatchParadigm=Batch, Model=Qwen3-4B2025.10 | 57.5 | 1,485.547 | 20.37 | |
| StreamingThinkerParadigm=Streaming, Reasoning Depth=D1, Model=Qwen3-4B2025.10 | 54.4 | 24.32 | 0.69 | |
| InterleavedParadigm=Interleaved, Reasoning Depth=D1, Model=Qwen3-4B2025.10 | 53.7 | 24.32 | 8.33 |