Next-action prediction on ScienceWorld
50.34AccuracyRL w/ ACT
Evaluation results
| Method | Links | |
|---|---|---|
| RL w/ ACTBackbone=Qwen3-8B2026.03 | 50.34 | |
| IL w/ ACTBackbone=Qwen3-8B2026.03 | 48.69 | |
| Early Experience (Self-Reflection)Backbone=Qwen3-8B2026.03 | 45.6 | |
| RLBackbone=Qwen3-8B2026.03 | 43.04 | |
| Imitation LearningBackbone=Qwen3-8B2026.03 | 42.8 | |
| Prompt w/o CoT thinkingBackbone=Qwen3-8B, CoT=false2026.03 | 28.01 | |
| ACTBackbone=Qwen3-8B2026.03 | 26.71 | |
| Prompt w/ CoT thinkingBackbone=Qwen3-8B, CoT=true2026.03 | 25.21 |