Multihop Question Answering on HotpotQA (EM, ASR)
75EMClean (No Attack)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Clean (No Attack)Framework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 75 | — | |
| TopicFramework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 75 | 1 | |
| CombinedFramework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 72 | 0 | |
| Fake CompletionFramework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 71 | 0 | |
| Clean (No Attack)Framework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 71 | — | |
| Fake CompletionFramework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 71 | 0 | |
| NaiveFramework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 70 | 1 | |
| IgnoreFramework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 70 | 1 | |
| CombinedFramework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 70 | 1 | |
| NaiveFramework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 69 | 6 | |
| NaiveFramework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 69 | 0 | |
| PARADOXFramework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 69 | 0 | |
| Clean (No Attack)Framework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 68 | — | |
| PARADOXFramework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 68 | 2 | |
| PARADOXFramework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 68 | 0 | |
| Clean (No Attack)Framework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 68 | — | |
| Fake CompletionFramework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 67 | 3 | |
| IgnoreFramework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 67 | 0 | |
| PoisonedRAGFramework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 66 | 0 | |
| PoisonedRAGFramework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 66 | 0 | |
| TopicFramework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 64 | 0 | |
| PoisonedRAGFramework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 58 | 0 | |
| NaiveFramework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 57 | 5 | |
| CombinedFramework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 56 | 3 | |
| IgnoreFramework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 55 | 6 | |
| PARADOXFramework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 55 | 2 | |
| KidnapRAGFramework=WebThinker, LLM Backbone=DeepSeek-32B2026.07 | 52 | 17 | |
| PoisonedRAGFramework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 50 | 0 | |
| IgnoreFramework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 40 | 21 | |
| TopicFramework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 40 | 0 | |
| Fake CompletionFramework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 33 | 38 | |
| CombinedFramework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 29 | 42 | |
| KidnapRAGFramework=WebThinker, LLM Backbone=QwQ-32B2026.07 | 23 | 60 | |
| KidnapRAGFramework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 20 | 39 | |
| TopicFramework=ReAct, LLM Backbone=Qwen2.5-32B-Inst2026.07 | 15 | 21 | |
| KidnapRAGFramework=ReAct, LLM Backbone=Llama-3.3-70B-Inst2026.07 | 13 | 47 |