Composing auto-generated skills on WebArena
57.6Skill Step PercentageGPT-5
Evaluation Results
| Method | Links | |
|---|---|---|
| GPT-5Evaluation Protocol=zero-shot format-constrained prompting2026.06 | 57.6 | |
| Claude Sonnet 4.5Evaluation Protocol=zero-shot format-constrained prompting2026.06 | 57.6 | |
| Llama-3.1-70BEvaluation Protocol=zero-shot format-constrained prompting2026.06 | 56.2 | |
| Qwen3-8BEvaluation Protocol=zero-shot2026.06 | 55.8 | |
| OLMo-3-7BEvaluation Protocol=zero-shot format-constrained prompting2026.06 | 54.5 | |
| Claude Haiku 4.5Evaluation Protocol=zero-shot format-constrained prompting2026.06 | 50.8 | |
| DeepSeek-R1-Distill-32BEvaluation Protocol=zero-shot format-constrained prompting2026.06 | 45.1 | |
| Qwen3-8BEvaluation Protocol=GRPO from base2026.06 | 44.2 | |
| Gemma-4-31BEvaluation Protocol=zero-shot format-constrained prompting2026.06 | 37.7 |