Tool-Augmented Agentic Reasoning on Full Evaluation Corpus (test)
2.99RewardSponge Tool Attack (STA)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Sponge Tool Attack (STA)Agentic Framework=OctoTools, Victim Model=Qwen2-VL-7B2026.01 | 2.99 | 47.27 | 45.58 | |
| Sponge Tool Attack (STA)Agentic Framework=LangChain, Victim Model=Qwen2-VL-7B2026.01 | 2.92 | 44.82 | 42.24 | |
| Sponge Tool Attack (STA)Agentic Framework=GPT-Functions, Victim Model=Qwen2-VL-7B2026.01 | 2.48 | 42.27 | 43.14 | |
| Sponge Tool Attack (STA)Agentic Framework=AutoGen, Victim Model=Qwen2-VL-7B2026.01 | 2.17 | 40.92 | 39.28 | |
| Sponge Tool Attack (STA)Agentic Framework=GPT-Functions, Victim Model=GPT-4o-mini2026.01 | 1.72 | 47.28 | 46.92 | |
| Sponge Tool Attack (STA)Agentic Framework=OctoTools, Victim Model=GPT-4o-mini2026.01 | 1.61 | 52.86 | 51.23 | |
| Sponge Tool Attack (STA)Agentic Framework=LangChain, Victim Model=GPT-4o-mini2026.01 | 1.49 | 48.15 | 47.62 | |
| Sponge Tool Attack (STA)Agentic Framework=AutoGen, Victim Model=GPT-4o-mini2026.01 | 1.35 | 45.73 | 43.24 |