Topic Classification on AGNews (Performance and Efficiency)
78.5Performance ScoreLLM w/o cascade
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| LLM w/o cascadeCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct2026.05 | 78.5 | 26.71 | 7.14 | 100 | |
| InitialCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct2026.05 | 78 | 17.38 | 4.73 | 62.5 | |
| Joint-target attackCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct, Setting=P=12026.05 | 70.5 | 24.15 | 6.59 | 85.5 | |
| Random noiseCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct2026.05 | 69.5 | 21.43 | 6.17 | 75.5 | |
| Single-target attackCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct, Setting=SLM only2026.05 | 68 | 20.09 | 5.73 | 70 | |
| Joint-target attackCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct, Setting=P=0.82026.05 | 61.5 | 20.97 | 5.84 | 72.5 | |
| Single-target attackCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct, Setting=DM only2026.05 | 56.5 | 15.17 | 4.38 | 55 | |
| Joint-target attackCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct, Setting=P=0.62026.05 | 53.5 | 17.24 | 4.76 | 61 | |
| Joint-target attackCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct, Setting=P=0.42026.05 | 44.5 | 13.85 | 4.01 | 47 | |
| Joint-target attackCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct, Setting=P=0.22026.05 | 37.5 | 9.64 | 3.15 | 31.5 | |
| Joint-target attackCascade system=GPT2-medium + BERT-base + Phi-3-medium-4k-instruct, Setting=P=02026.05 | 19.5 | 5.78 | 2.25 | 18 |