List your company
ResearchFormerly WizWand

Multi-hop Question Answering on 2WikiMultiHopQA (val)

69.3Exact Match (EM)GRPO
-0.58817.55635.753.844Feb 26, 2026Feb 28, 2026Mar 3, 2026Mar 6, 2026Mar 8, 2026Mar 11, 2026Mar 14, 2026
Updated 5mo ago

Evaluation results

MethodLinks
GRPOModel Size=Qwen3 4B2026.0269.3————77.7387.53
MBT-RModel Size=Qwen3 4B2026.0268.26————77.9287.25
MBT-SModel Size=Qwen3 4B2026.0267.52————77.3286.67
GRPOModel Size=Qwen3 1.7B2026.0263.25————72.3280.69
MBT-SModel Size=Qwen3 1.7B2026.0262.85————72.3780.66
MBT-RModel Size=Qwen3 1.7B2026.0262.24————72.2280.62
RSModel Size=Qwen3 1.7B2026.0258.06————68.2676.75
MBT-SModel Size=Qwen3 0.6B2026.0257.78————67.0573.32
MBT-RModel Size=Qwen3 0.6B2026.0257.21————66.3573.17
TokenSkipModel Size=Qwen3 1.7B2026.0257————66.3774.65
GRPOModel Size=Qwen3 0.6B2026.0254.61————62.9369.25
Apex-SearcherType=Agentic RAG, Backbone=Qwen2.5-7B-Instruct2026.0354——————
BaseModel Size=Qwen3 1.7B2026.0252.23————61.6569.59
BaseModel Size=Qwen3 4B2026.0250.7————62.0786.43
RSModel Size=Qwen3 4B2026.0250.3————61.5985.53
DecEx-RAGType=Agentic RAG, Backbone=Qwen2.5-7B-Instruct2026.0350——————
Apex-SearcherType=Agentic RAG, Backbone=Qwen2.5-3B-Instruct2026.0349.4——————
TokenSkipModel Size=Qwen3 4B2026.0249.36————59.8783.42
PromptModel Size=Qwen3 4B2026.0249.2————60.2186.27
ReasonRAGType=Agentic RAG, Backbone=Qwen2.5-7B-Instruct2026.0343.6——————
PromptModel Size=Qwen3 1.7B2026.0242.46————55.6374.19
Search-R1-InstructType=Agentic RAG, Backbone=Qwen2.5-7B-Instruct2026.0341.4——————
LIMOProModel Size=Qwen3 4B2026.0241.01————56.2980.57
LIMOProModel Size=Qwen3 1.7B2026.0239.15————54.776.43
StepSearch-instructType=Agentic RAG, Backbone=Qwen2.5-7B-Instruct2026.0336.6——————
RSModel Size=Qwen3 0.6B2026.0236.23————47.0455.42
ZeroSearch-instructType=Agentic RAG, Backbone=Qwen2.5-7B-Instruct2026.0335.2——————
BaseModel Size=Qwen3 0.6B2026.0233.99————43.951.59
TokenSkipModel Size=Qwen3 0.6B2026.0233.95————44.552.81
LIMOProModel Size=Qwen3 0.6B2026.0232.88————48.368.39
StepSearch-instructType=Agentic RAG, Backbone=Qwen2.5-3B-Instruct2026.0332——————
ZeroSearch-instructType=Agentic RAG, Backbone=Qwen2.5-3B-Instruct2026.0330——————
PromptModel Size=Qwen3 0.6B2026.0229.64————39.6349.84
Search-R1-InstructType=Agentic RAG, Backbone=Qwen2.5-3B-Instruct2026.0327.3——————
Direct InferenceType=Non-Retrieval, Backbone=Qwen2.5-7B-Instruct2026.0325——————
Direct InferenceType=Non-Retrieval, Backbone=Qwen2.5-3B-Instruct2026.0324.4——————
RAGType=Standard RAG, Backbone=Qwen2.5-7B-Instruct2026.0323.5——————
RAGType=Standard RAG, Backbone=Qwen2.5-3B-Instruct2026.0322.6——————
Search-o1Type=Agentic RAG, Backbone=Qwen2.5-3B-Instruct2026.0321.8——————
Search-o1Type=Agentic RAG, Backbone=Qwen2.5-7B-Instruct2026.0317.6——————
IRCoTType=Iterative RAG, Backbone=Qwen2.5-3B-Instruct2026.0317.1——————
IRCoTType=Iterative RAG, Backbone=Qwen2.5-7B-Instruct2026.0314.9——————
CoTType=Non-Retrieval, Backbone=Qwen2.5-7B-Instruct2026.0311.1——————
CoTType=Non-Retrieval, Backbone=Qwen2.5-3B-Instruct2026.032.1——————
Base + toolsAugmentation=Information retrieval tool, Protocol=Agentic RAG pipeline2026.02—23726————
Base Qwen3-4BContext=Full context2026.02—651,214————
LogicPoisonLLM=GPT-4o-mini, RAG Framework=Naive RAG2026.04———9091.6——
LogicPoisonLLM=GPT-4o-mini, RAG Framework=GraphRAG2026.04———78.495.6——
LogicPoisonLLM=GPT-4o-mini, RAG Framework=GFM-RAG2026.04———71.676——
LogicPoisonLLM=GPT-4o-mini, RAG Framework=HippoRAG22026.04———82.871.8——
LogicPoisonLLM=Llama-3.1-8B, RAG Framework=Naive RAG2026.04———95.494.4——
LogicPoisonLLM=Llama-3.1-8B, RAG Framework=GraphRAG2026.04———78.895——
LogicPoisonLLM=Llama-3.1-8B, RAG Framework=GFM-RAG2026.04———74.285.6——
LogicPoisonLLM=Llama-3.1-8B, RAG Framework=HippoRAG22026.04———77.474.2——
LogicPoisonLLM=Qwen-3-32B, RAG Framework=Naive RAG2026.04———82.283.4——
LogicPoisonLLM=Qwen-3-32B, RAG Framework=GraphRAG2026.04———78.693.2——
LogicPoisonLLM=Qwen-3-32B, RAG Framework=GFM-RAG2026.04———71.479.8——
LogicPoisonLLM=Qwen-3-32B, RAG Framework=HippoRAG22026.04———87.476.6——
PoisonedRAGLLM=GPT-4o-mini, RAG Framework=Naive RAG2026.04———64.483.4——
PoisonedRAGLLM=GPT-4o-mini, RAG Framework=GraphRAG2026.04———58.677.4——
PoisonedRAGLLM=GPT-4o-mini, RAG Framework=GFM-RAG2026.04———57.256.6——
PoisonedRAGLLM=GPT-4o-mini, RAG Framework=HippoRAG22026.04———74.870.4——
PoisonedRAGLLM=Llama-3.1-8B, RAG Framework=Naive RAG2026.04———61.477.8——
PoisonedRAGLLM=Llama-3.1-8B, RAG Framework=GraphRAG2026.04———54.873.2——
PoisonedRAGLLM=Llama-3.1-8B, RAG Framework=GFM-RAG2026.04———42.237.2——
PoisonedRAGLLM=Llama-3.1-8B, RAG Framework=HippoRAG22026.04———74.870——
PoisonedRAGLLM=Qwen-3-32B, RAG Framework=Naive RAG2026.04———6382.4——
PoisonedRAGLLM=Qwen-3-32B, RAG Framework=GraphRAG2026.04———5877——
PoisonedRAGLLM=Qwen-3-32B, RAG Framework=GFM-RAG2026.04———58.837.6——
PoisonedRAGLLM=Qwen-3-32B, RAG Framework=HippoRAG22026.04———75.873.6——
SELFCESTMechanism=Spawn clones with retrieval key, Training=End-to-end reinforcement learning2026.02—70644————