Question Answering on CrossThink-QA
50.2AccuracyFlowTracer
Evaluation Results
| Method | Links | |
|---|---|---|
| FlowTracerBackbone=Qwen3-4B-Base, Context window=1K2026.06 | 50.2 | |
| GRPO + coupled rhythm creditBackbone=Qwen3-4B-Base2025.10 | 50.1 | |
| GRPO + local-chunk creditBackbone=Qwen3-4B-Base2025.10 | 50 | |
| GRPO + global-anchor creditBackbone=Qwen3-4B-Base2025.10 | 49.6 | |
| AttentionBackbone=Qwen3-4B-Base, Context window=1K2026.06 | 49.6 | |
| GRPOBackbone=Qwen3-4B-Base2025.10 | 48 | |
| GRPO + high-entropy creditBackbone=Qwen3-4B-Base2025.10 | 48 | |
| GRPOBackbone=Qwen3-4B-Base, Context window=1K2026.06 | 48 | |
| GRPO + random creditBackbone=Qwen3-4B-Base2025.10 | 47.8 | |
| RandomBackbone=Qwen3-4B-Base, Context window=1K2026.06 | 47.6 | |
| High-entropyBackbone=Qwen3-4B-Base, Context window=1K2026.06 | 47.6 |