List your company

Multi-hop Question Answering on MuSiQue (val)

810Avg Token BudgetSELFCEST
651.9281,718.9142,785.93,852.886Feb 3, 2026Feb 28, 2026Mar 26, 2026Apr 21, 2026May 16, 2026Jun 11, 2026Jul 7, 2026

Evaluation results

MethodLinks
SELFCESTMechanism=Spawn clones with retrieval key, Training=End-to-end reinforcement learning2026.0281051———
Base + toolsAugmentation=Information retrieval tool, Protocol=Agentic RAG pipeline2026.0286014———
CRAGBackbone=Llama-3.1-8B2026.071,477.9—19.8—10.76
Adaptive-RAGBackbone=Llama-3.1-8B2026.071,749.8—17.19—9.52
Adaptive-RAGBackbone=GPT-4o-mini2026.071,789.1—31.72—19.28
IRCoTBackbone=GPT-4o-mini2026.071,939.6—38.7—24.53
CRAGBackbone=GPT-4o-mini2026.072,020.7—35.78—21.51
IRCoTBackbone=Llama-3.1-8B2026.072,049.1—21.26—13.16
IRCoTBackbone=Qwen2.5-7B2026.072,109.6—28.89—18.62
Adaptive-RAGBackbone=Qwen2.5-7B2026.072,160.4—25.79—16.47
Base Qwen3-4BContext=Full context2026.022,41735———
CRAGBackbone=Qwen2.5-7B2026.072,479.5—27.4—16.76
DynaKRAGBackbone=GPT-4o-mini2026.072,499.9—39.77—26.35
Fixed-KBackbone=GPT-4o-mini2026.072,763.6—32.79—18.99
Fixed-KBackbone=Llama-3.1-8B2026.072,814.2—14.61—6.95
Fixed-KBackbone=Qwen2.5-7B2026.072,918.3—25.53—15.18
CoRAG-s3Backbone=Llama-3.1-8B2026.072,945.1—16.9—9.56
CoRAG-s3Backbone=GPT-4o-mini2026.073,046—38.72—24.82
DynaKRAGBackbone=Qwen2.5-7B2026.073,082.3—30.61—19.86
S2G-RAGBackbone=GPT-4o-mini2026.073,091.2—36.91—22.8
DynaKRAGBackbone=Llama-3.1-8B2026.073,254.6—23.91—15.14
CoRAG-s3Backbone=Qwen2.5-7B2026.073,263.6—28.48—18.58
PAR2-RAGBackbone=Llama-3.1-8B2026.073,333.6—16.74—8.81
PAR2-RAGBackbone=GPT-4o-mini2026.073,477.4—32.41—18.87
PAR2-RAGBackbone=Qwen2.5-7B2026.073,755.6—24.8—14.48
S2G-RAGBackbone=Llama-3.1-8B2026.073,848.5—19.73—11.46
S2G-RAGBackbone=Qwen2.5-7B2026.073,886.8—29.99—19.36
Self-Ask+SearchBackbone=Llama-3.1-8B2026.074,000.1—16.85—9.35
Self-Ask+SearchBackbone=GPT-4o-mini2026.074,570—38.76—24.49
Self-Ask+SearchBackbone=Qwen2.5-7B2026.074,761.8—29.98—19.74
Qwen3-32B-BaseSetting=Zero-shot, Mode=Closed-book2026.05—3.1313.14——
Qwen3-32B-DS-SFTSetting=Zero-shot, Mode=Closed-book, Teacher LLM=DeepSeek2026.05—3.413.6119.7—
Qwen3-32B-GE-SFTSetting=Zero-shot, Mode=Closed-book, Teacher LLM=Gemini2026.05—3.213.6513.6—