Agent Task on ToolBench
44.98Success RatePAIR
Evaluation Results
| Method | Links | |
|---|---|---|
| PAIR2026.05 | 44.98 | |
| LLM-as-a-judge2026.05 | 42.03 | |
| CoE-C2026.05 | 40.34 | |
| IGPO2026.05 | 39.52 | |
| Tree-GRPO2026.05 | 39.35 | |
| Lookback-ratio2026.05 | 39.34 | |
| Mean-pooled2026.05 | 39.17 | |
| Outcome2026.05 | 38.78 | |
| Head-entropy2026.05 | 38.68 | |
| Multi-layer2026.05 | 38.56 | |
| Last-token2026.05 | 38.37 | |
| AT2PO2026.05 | 38.12 | |
| Hidden+Attn2026.05 | 37.84 | |
| Multi-Attn2026.05 | 36.14 | |
| TIPS2026.05 | 34.83 | |
| Attention2026.05 | 34.71 |