List your company

Multi-Hop Search-augmented Question Answering on HotpotQA (success rate (%))

38.8Success RateRole-Agent
23.09627.17331.2535.327May 29, 2026May 30, 2026Jun 1, 2026Jun 3, 2026Jun 5, 2026Jun 7, 2026Jun 9, 2026

Evaluation results

MethodLinks
Role-AgentType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.0638.8
GiGPOType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.0636.9
StepSearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.0634.5
+ MASABackbone=Qwen3-32B2026.0534.2
+ DS-AdapterBackbone=Qwen3-32B2026.0534
+ Base SkillBackbone=Qwen3-32B2026.0533.8
+ MASABackbone=Qwen3-14B2026.0532.8
+ Base SkillBackbone=Qwen3-14B2026.0532.7
Search_R1Type=RL Training, Backbone=Qwen2.5-3B-Instruct2026.0632.4
No SkillBackbone=Qwen3-32B2026.0532.2
No SkillBackbone=Qwen3-14B2026.0531.7
+ DS-AdapterBackbone=Qwen3-14B2026.0531.6
+ DS-AdapterBackbone=Qwen3-4B2026.0528.6
+ Base SkillBackbone=Qwen3-8B2026.0528.6
+ MASABackbone=Qwen3-8B2026.0528.6
+ Base SkillBackbone=Qwen3-4B2026.0528.5
+ DS-AdapterBackbone=Qwen3-8B2026.0527.8
No SkillBackbone=Qwen3-4B2026.0527.7
+ MASABackbone=Qwen3-4B2026.0527.4
Zero-SearchType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.0627.4
No SkillBackbone=Qwen3-8B2026.0524.8
R1-InstructType=RL Training, Backbone=Qwen2.5-3B-Instruct2026.0623.7