Lifelong Agent Interaction on SciWorld
73.5Success Rate (SR)PROACTAGENT
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| PROACTAGENTRegime=Ours2026.04 | 73.5 | 18.38 | |
| GRPO+ReflexionRegime=Online baselines2026.04 | 55.5 | 27.52 | |
| GRPORegime=Offline baselines2026.04 | 52 | 26.51 | |
| GRPO (online)Regime=Online baselines2026.04 | 46.5 | 27.87 | |
| SFTRegime=Offline baselines2026.04 | 19.5 | 30.26 | |
| ReflexionRegime=Online baselines2026.04 | 4 | 33.73 | |
| Mem0Regime=Online baselines2026.04 | 3 | 34.94 | |
| MemoryBankRegime=Online baselines2026.04 | 2.5 | 30.24 | |
| BaselineRegime=Offline baselines2026.04 | 2 | 33.4 | |
| AWMRegime=Online baselines2026.04 | 1 | 35.65 |