Interactive Task Completion on ScienceWorld
67.2L0 ScoreRLVMR
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| RLVMRModel Backbone=Qwen2.5-7B-Instruct2026.04 | 67.2 | 43 | 32.2 | 47.5 | |
| DPEPOModel Backbone=Qwen2.5-7B-Instruct2026.04 | 66.6 | 66.5 | 51 | 61.4 | |
| ETOModel Backbone=Qwen2.5-7B-Instruct2026.04 | 62.5 | 40.6 | 28.1 | 43.7 | |
| DPEPOModel Backbone=Qwen2.5-1.5B-Instruct2026.04 | 59.8 | 58.1 | 34.2 | 50.7 | |
| GiGPOModel Backbone=Qwen2.5-7B-Instruct2026.04 | 53.4 | 25.2 | 25.8 | 34.8 | |
| GRPOModel Backbone=Qwen2.5-7B-Instruct2026.04 | 49.1 | 30.1 | 26.6 | 35.3 | |
| RLVMRModel Backbone=Qwen2.5-1.5B-Instruct2026.04 | 46.9 | 34.3 | 26.5 | 35.9 | |
| GPT-4oModel Backbone=Frontier LLMs2026.04 | 45.4 | 49.2 | 41 | 45.2 | |
| ETOModel Backbone=Qwen2.5-1.5B-Instruct2026.04 | 39.1 | 22.7 | 15.6 | 25.8 | |
| DeepSeek-V3Model Backbone=Frontier LLMs2026.04 | 27.3 | 35.2 | 26.5 | 29.7 | |
| GiGPOModel Backbone=Qwen2.5-1.5B-Instruct2026.04 | 25.8 | 15.2 | 4.7 | 15.2 | |
| DeepSeek-R1Model Backbone=Frontier LLMs2026.04 | 22.2 | 31.4 | 29.1 | 27.6 | |
| GRPOModel Backbone=Qwen2.5-1.5B-Instruct2026.04 | 21.1 | 13.7 | 10.9 | 15.2 |