List your company

Multi-hop Question Answering on MuSiQue (OOD)

20.6EMRECON
1.06886.139411.2116.2806Jul 21, 2025Aug 3, 2025Aug 17, 2025Aug 31, 2025Sep 14, 2025Sep 28, 2025Oct 12, 2025

Evaluation results

MethodLinks
RECONBackbone=Qwen2.5-7B-Base, Training Protocol=PPO, RECON Stage=Stage 1 + Stage 22025.1020.6—
Search-R1Backbone=Qwen2.5-7B-Base, Training Protocol=PPO2025.1019.6—
EviOmniMethod Category=RAR, Compression Ratio=96.60x2025.071855.5
CoTMethod Category=RAR, Compression Ratio=11.25x2025.0716.0153.05
RECONBackbone=Qwen2.5-7B-Base, Training Protocol=PPO, RECON Stage=Stage 2 Distillation-only2025.1012.8—
Rejection Sampling2025.1012.3—
FullMethod Category=WR, Compression Ratio=1.0x2025.0711.7548.89
FilCoMethod Category=VAR, Compression Ratio=67.50x2025.078.3645.74
R1-base2025.108.3—
RECONBackbone=Qwen2.5-3B-Base, Training Protocol=PPO, RECON Stage=Stage 1 + Stage 22025.108.2—
LLMLingua-2Method Category=VAR, Compression Ratio=4.55x2025.077.745.46
IRCoT2025.107.2—
R1-instruct2025.107.2—
SFT2025.106.6—
Search-o12025.105.8—
RAG2025.105.8—
RECONBackbone=Qwen2.5-3B-Base, Training Protocol=PPO, RECON Stage=Stage 2 Distillation-only2025.105.7—
Search-R1Backbone=Qwen2.5-3B-Base, Training Protocol=PPO2025.104.9—
Direct Inference2025.103.1—
CoT2025.102.2—
ZeroMethod Category=WR2025.071.8241.92