List your company

Multi-hop Question Answering on 2WikiMultiHopQA (EM, Tok)

48.3EMCARL
18.24426.04733.8541.653May 27, 2026Jun 2, 2026Jun 9, 2026Jun 16, 2026Jun 23, 2026Jun 30, 2026Jul 7, 2026

Evaluation results

MethodLinks
CARLModel Scale=7B-Inst.2026.0548.31,139
CARLModel Scale=3B-Inst.2026.0544.71,394
IGRPOBackbone=Qwen2.5-7B-Instruct, Type=Tree-based2026.0743.8—
GiGPOBackbone=Qwen2.5-7B-Instruct, Type=Chain-based2026.0743.6—
IGPOBackbone=Qwen2.5-7B-Instruct, Type=Chain-based2026.0742.5—
Search-R1 PPOModel Scale=7B-Inst.2026.0541.71,775
IGRPOBackbone=Qwen2.5-3B-Instruct, Type=Tree-based2026.0740.2—
Search-R1Backbone=Qwen2.5-7B-Instruct, Type=Chain-based2026.0740.1—
GiGPOBackbone=Qwen2.5-3B-Instruct, Type=Chain-based2026.0737—
Search-R1 GRPOModel Scale=7B-Inst.2026.0536.81,827
StepSearchBackbone=Qwen2.5-7B-Instruct, Type=Chain-based2026.0736.6—
Tree-GRPOBackbone=Qwen2.5-7B-Instruct, Type=Tree-based2026.0735.6—
ZeroSearchBackbone=Qwen2.5-7B-Instruct, Type=Chain-based2026.0735.2—
IGPOBackbone=Qwen2.5-3B-Instruct, Type=Chain-based2026.0734.9—
Search-R1 GRPOModel Scale=3B-Inst.2026.0534.31,960
SFT (rej. sampled)Model Scale=7B-Inst.2026.0532.61,683
Search-R1 PPOModel Scale=3B-Inst.2026.0532.11,861
StepSearchBackbone=Qwen2.5-3B-Instruct, Type=Chain-based2026.0732—
Search-R1Backbone=Qwen2.5-3B-Instruct, Type=Chain-based2026.0731.9—
Tree-GRPOBackbone=Qwen2.5-3B-Instruct, Type=Tree-based2026.0731.7—
ZeroSearchBackbone=Qwen2.5-3B-Instruct, Type=Chain-based2026.0730—
R1-InstructBackbone=Qwen2.5-3B-Instruct, Type=RL Training2026.0729.2—
R1-InstructBackbone=Qwen2.5-7B-Instruct, Type=RL Training2026.0727.5—
CoT+Tools (selective)Model Scale=7B-Inst.2026.0527.11,641
CoT+Tools (always)Model Scale=7B-Inst.2026.0526.61,938
SFT (rej. sampled)Model Scale=3B-Inst.2026.0525.11,791
CoT+Tools (selective)Model Scale=3B-Inst.2026.0521.21,732
CoT+Tools (always)Model Scale=3B-Inst.2026.0519.42,028