Node Count on synthetic benchmark 1.0 (test)
100AccuracyQwen3-LoRA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Qwen3-LoRABackbone=Qwen3-600M, Training=LoRA fine-tuning2026.01 | 100 | 0 | |
| NAG-LoRABackbone=Qwen3-600M, Training=LoRA fine-tuning2026.01 | 100 | 0 | |
| NAG-ZeroBackbone=Qwen3-600M, Mode=Zero-shot2026.01 | 99.85 | 0 | |
| GraphTokenNBackbone=Qwen3-600M, Granularity=Node-level embeddings2026.01 | 94.38 | 0.06 | |
| GraphTokenEBackbone=Qwen3-600M, Granularity=Edge-level embeddings2026.01 | 93.26 | 0.13 | |
| GraphTokenGBackbone=Qwen3-600M, Granularity=Global pooled representation2026.01 | 82.02 | 0.19 | |
| Qwen3-DirectBackbone=Qwen3-8B, Mode=Zero-shot2026.01 | 45.65 | 3.81 |