List your company

Multi-hop Question Answering on HotpotQA (Recall@K and NDCG@K)

48.1Recall@1GRPO
-1.61211.29424.237.106May 13, 2026

Evaluation results

MethodLinks
GRPOBackbone=Qwen3.5-2B2026.0548.186.386.596.188.488.5
F-GRPOBackbone=Qwen3.5-2B2026.0548.191.491.496.392.592.5
GRPOBackbone=Qwen3-4B2026.05489393.196.19393.1
F-GRPOBackbone=Qwen3-4B2026.054892.692.69693.393.3
RankZephyr2026.0547.484.990.394.885.988.6
MonoT52026.0546.386.292.892.586.289.5
LiT52026.0546.378.485.292.580.383.7
Decoupled SFT (rank. only)Backbone=Qwen3-4B2026.0543.884.389.287.68486.5
DuoT52026.0543.279.886.286.48083.3
BM252026.0541.872.282.783.673.278.5
SFTBackbone=Qwen3-4B2026.0538.368.77376.669.671.8
Decoupled SFT (full)Backbone=Qwen3-4B2026.0532.64952.965.15254
Zero-shot (single-step)Backbone=Qwen3-4B2026.0530.76165.461.460.262.4
Decoupled SFT (rank. only)Backbone=Qwen3.5-2B2026.0521.451.768.542.847.756.1
Decoupled SFT (sel. only)Backbone=Qwen3-4B2026.059.819.120.619.618.919.7
Decoupled SFT (full)Backbone=Qwen3.5-2B2026.05616.625.711.914.819.3
SFTBackbone=Qwen3.5-2B2026.055.215.125.310.413.218.2
Random2026.0551525.19.913.118.1
Zero-shot (single-step)Backbone=Qwen3.5-2B2026.0547.48.487.47.9
Zero-shot (two-step)Backbone=Qwen3-4B2026.0535.86.165.85.9
Zero-shot (two-step)Backbone=Qwen3.5-2B2026.0511.92.121.92
Decoupled SFT (sel. only)Backbone=Qwen3.5-2B2026.050.30.60.70.50.50.6