Planning-layer poisoning attack on DeepResearch Bench stratified 10-query subset 1.0
38.5PRISMFORGE
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| FORGENetwork j=5, Document diversity=true, Intra-doc reasoning=true, Inter-doc chain=true2026.07 | 38.5 | 37.7 | 22.5 | 33.1 | 44.5 | 49.5 | |
| B2 (FORGE ablation)Network j=5, Document diversity=true, Intra-doc reasoning=true, Inter-doc chain=false2026.07 | 27.9 | 37.8 | 12.8 | 25.3 | 29.9 | 28.4 | |
| PoisonedRAGNetwork j=52026.07 | 23 | 11.2 | 7.3 | 21.7 | 42.8 | 28.2 | |
| B1 (FORGE ablation)Network j=5, Document diversity=true, Intra-doc reasoning=false, Inter-doc chain=false2026.07 | 20.6 | 21.2 | 13.8 | 24.8 | 17.8 | 33.7 | |
| AuthChainNetwork j=52026.07 | 18.8 | 5.8 | 5.8 | 31 | 21.2 | 33.2 | |
| No AttackNetwork j=52026.07 | 0 | — | — | — | — | — |