Embodied Agent Task on ScienceWorld Unseen
70.8Success RateGRPO w/ EVU
Evaluation Results
| Method | Links | |
|---|---|---|
| GRPO w/ EVUMethod Category=Training, Backbone Model=Qwen2.5-3B-Instruct, Base Method=GRPO, EVU Mechanism=Included2026.04 | 70.8 | |
| PPO w/ EVUMethod Category=Training, Backbone Model=Qwen2.5-3B-Instruct, Base Method=PPO, EVU Mechanism=Included2026.04 | 62.7 | |
| ReAct w/ EVUMethod Category=Prompting, Backbone Model=DeepSeek V3.2, Base Method=ReAct, EVU Mechanism=Included2026.04 | 60.9 | |
| ReActMethod Category=Prompting, Backbone Model=DeepSeek V3.2, Base Method=ReAct, EVU Mechanism=None2026.04 | 57.8 | |
| Plan-and-Act w/ EVUMethod Category=Prompting, Backbone Model=DeepSeek V3.2, Base Method=Plan-and-Act, EVU Mechanism=Included2026.04 | 55.3 | |
| PPO w/ EVUMethod Category=Training, Backbone Model=Qwen3-1.7B-Instruct, Base Method=PPO, EVU Mechanism=Included2026.04 | 54 | |
| GRPOMethod Category=Training, Backbone Model=Qwen2.5-3B-Instruct, Base Method=GRPO, EVU Mechanism=None2026.04 | 52.8 | |
| NoThinking w/ EVUMethod Category=Prompting, Backbone Model=DeepSeek V3.2, Base Method=NoThinking, EVU Mechanism=Included2026.04 | 52.2 | |
| PPOMethod Category=Training, Backbone Model=Qwen2.5-3B-Instruct, Base Method=PPO, EVU Mechanism=None2026.04 | 51.6 | |
| Plan-and-ActMethod Category=Prompting, Backbone Model=DeepSeek V3.2, Base Method=Plan-and-Act, EVU Mechanism=None2026.04 | 50.9 | |
| GRPO w/ EVUMethod Category=Training, Backbone Model=Qwen3-1.7B-Instruct, Base Method=GRPO, EVU Mechanism=Included2026.04 | 50.3 | |
| NoThinkingMethod Category=Prompting, Backbone Model=DeepSeek V3.2, Base Method=NoThinking, EVU Mechanism=None2026.04 | 46 | |
| SFT w/ EVUMethod Category=Training, Backbone Model=Qwen2.5-3B-Instruct, Base Method=SFT, EVU Mechanism=Included2026.04 | 45.3 | |
| GRPOMethod Category=Training, Backbone Model=Qwen3-1.7B-Instruct, Base Method=GRPO, EVU Mechanism=None2026.04 | 42.9 | |
| PPOMethod Category=Training, Backbone Model=Qwen3-1.7B-Instruct, Base Method=PPO, EVU Mechanism=None2026.04 | 41 | |
| SFT w/ EVUMethod Category=Training, Backbone Model=Qwen3-1.7B-Instruct, Base Method=SFT, EVU Mechanism=Included2026.04 | 24.8 | |
| SFTMethod Category=Training, Backbone Model=Qwen2.5-3B-Instruct, Base Method=SFT, EVU Mechanism=None2026.04 | 13.7 | |
| SFTMethod Category=Training, Backbone Model=Qwen3-1.7B-Instruct, Base Method=SFT, EVU Mechanism=None2026.04 | 11.2 |