Web Agent Task Success on WebArena
27.8Task Success Rate (TSR)DPO – ADARUBRIC-DA
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DPO – ADARUBRIC-DABackbone=Qwen2.5-7B, Evaluation Protocol=DPO training2026.03 | 27.8 | — | |
| DPO – ADARUBRIC-GMBackbone=Qwen2.5-7B, Evaluation Protocol=DPO training2026.03 | 25.1 | — | |
| DPO – ADARUBRIC-WMBackbone=Qwen2.5-7B, Evaluation Protocol=DPO training2026.03 | 24.3 | — | |
| DPO – PrometheusBackbone=Qwen2.5-7B, Evaluation Protocol=DPO training2026.03 | 21 | — | |
| DPO – G-EvalBackbone=Qwen2.5-7B, Evaluation Protocol=DPO training2026.03 | 20.1 | — | |
| DPO – Random pairsBackbone=Qwen2.5-7B, Evaluation Protocol=DPO training2026.03 | 17.4 | — | |
| Qwen 2.5 7B Instruct (Environment Poisoning)Defense=None, Stage=SFT, Poison %=252025.10 | 17 | 100 | |
| SFT – Success onlyBackbone=Qwen2.5-7B2026.03 | 16.7 | — | |
| Qwen 2.5 7B Instruct (Environment Poisoning)Defense=Granite Guardian 3.3-8B (GG), Stage=SFT, Poison %=252025.10 | 16.27 | 95.15 | |
| Qwen 2.5 7B Instruct (Environment Poisoning)Defense=Llama-firewall (LF), Stage=SFT, Poison %=252025.10 | 15.76 | 93.94 | |
| BaseBackbone=Qwen2.5-7B, Evaluation Protocol=zero-shot2026.03 | 12.3 | — | |
| Qwen 2.5 7B InstructDefense=None, Stage=Baseline, Poison %=02025.10 | 0.6 | — |