Interactive LLM Agent Task Success on SciWorld Unseen
82.4Success RateQ-Evolve
Evaluation results
| Method | Links | |
|---|---|---|
| Q-EvolveBackbone=Llama-3-8B-Instruct2026.06 | 82.4 | |
| ETO + MPOBackbone=Llama-3-8B-Instruct2026.06 | 80.8 | |
| WKMBackbone=Llama-3-8B-Instruct2026.06 | 76.5 | |
| ETOBackbone=Llama-3-8B-Instruct2026.06 | 74.1 | |
| KnowAgentBackbone=Llama-3-8B-Instruct2026.06 | 69.6 | |
| SFT + MPOBackbone=Llama-3-8B-Instruct2026.06 | 65.9 | |
| SFTBackbone=Llama-3-8B-Instruct2026.06 | 57 |