Commonsense Reasoning on CSQA
96AccuracyToken-ICS
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Token-ICSModel=Qwen-32B2026.02 | 96 | — | — | — | — | — | |
| Thought-ICSModel=Qwen-32B2026.02 | 96 | — | — | — | — | — | |
| Thought-ICSModel=Qwen-7B2026.02 | 94 | — | — | — | — | — | |
| Token-ICSModel=Qwen-14B2026.02 | 94 | — | — | — | — | — | |
| Thought-ICSModel=LLaMA-70B2026.02 | 94 | — | — | — | — | — | |
| Thought-ICSModel=OSS-120B2026.02 | 94 | — | — | — | — | — | |
| Thought-ICSModel=LLaMA-8B2026.02 | 93 | — | — | — | — | — | |
| Thought-ICSModel=Qwen-14B2026.02 | 93 | — | — | — | — | — | |
| Thought-ICSModel=OSS-20B2026.02 | 93 | — | — | — | — | — | |
| MV@10Model=Qwen-32B2026.02 | 92 | — | — | — | — | — | |
| Token-ICSModel=OSS-120B2026.02 | 92 | — | — | — | — | — | |
| BaselineMAS Topology=Star2025.11 | 91.75 | — | — | — | — | — | |
| BaselineMAS Topology=chain2025.11 | 91.46 | — | — | — | — | — | |
| BaselineMAS Topology=complete2025.11 | 90.93 | — | — | — | — | — | |
| MAS-ShieldMAS Topology=Star2025.11 | 90.78 | — | — | — | — | — | |
| BaselineMAS Topology=tree2025.11 | 90.68 | — | — | — | — | — | |
| MAS-ShieldMAS Topology=chain2025.11 | 90.64 | — | — | — | — | — | |
| MAS-ShieldMAS Topology=complete2025.11 | 90.13 | — | — | — | — | — | |
| Thought-ICS-SModel=Qwen-14B2026.02 | 90 | — | — | — | — | — | |
| Thought-ICS-AModel=Qwen-14B2026.02 | 90 | — | — | — | — | — | |
| Thought-ICS-SModel=Qwen-32B2026.02 | 90 | — | — | — | — | — | |
| Thought-ICS-AModel=Qwen-32B2026.02 | 90 | — | — | — | — | — | |
| Token-ICSModel=LLaMA-70B2026.02 | 90 | — | — | — | — | — | |
| MV@10Model=OSS-120B2026.02 | 90 | — | — | — | — | — | |
| MAS-ShieldMAS Topology=tree2025.11 | 89.31 | — | — | — | — | — | |
| CoTModel=Qwen-32B2026.02 | 89 | — | — | — | — | — | |
| InitModel=Qwen-32B2026.02 | 89 | — | — | — | — | — | |
| MV@10Model=LLaMA-70B2026.02 | 89 | — | — | — | — | — | |
| BaselineMAS Topology=cycle2025.11 | 88.91 | — | — | — | — | — | |
| Self-RefineModel=Qwen-32B2026.02 | 88 | — | — | — | — | — | |
| Thought-ICSModel=LLaMA-3B2026.02 | 88 | — | — | — | — | — | |
| MV@10Model=Qwen-14B2026.02 | 88 | — | — | — | — | — | |
| MAS-ShieldMAS Topology=cycle2025.11 | 87.55 | — | — | — | — | — | |
| Thought-ICS-SModel=LLaMA-70B2026.02 | 87 | — | — | — | — | — | |
| Thought-ICS-AModel=LLaMA-70B2026.02 | 87 | — | — | — | — | — | |
| MV@10Model=Qwen-7B2026.02 | 87 | — | — | — | — | — | |
| CoVeModel=Qwen-32B2026.02 | 86 | — | — | — | — | — | |
| Thought-ICS-SModel=Qwen-7B2026.02 | 85 | — | — | — | — | — | |
| Thought-ICS-AModel=Qwen-7B2026.02 | 85 | — | — | — | — | — | |
| Self-RefineModel=Qwen-14B2026.02 | 85 | — | — | — | — | — | |
| CoTModel=LLaMA-70B2026.02 | 85 | — | — | — | — | — | |
| CoVeModel=OSS-120B2026.02 | 85 | — | — | — | — | — | |
| InitModel=LLaMA-70B2026.02 | 85 | — | — | — | — | — | |
| CoTModel=Qwen-14B2026.02 | 84 | — | — | — | — | — | |
| Self-RefineModel=LLaMA-70B2026.02 | 84 | — | — | — | — | — | |
| CoTModel=OSS-120B2026.02 | 84 | — | — | — | — | — | |
| InitModel=Qwen-14B2026.02 | 84 | — | — | — | — | — | |
| InitModel=OSS-120B2026.02 | 84 | — | — | — | — | — | |
| MINDBackbone=Qwen2.5-7B, Fusion=with2026.01 | 83.98 | — | — | — | — | — | |
| EDITStudent Architecture=Qwen2.5-7B, Distillation Method=EDIT2026.01 | 83.8 | — | — | — | — | — | |
| EDITBackbone=Qwen2.5-7B, Fusion=none2026.01 | 83.8 | — | 5.13 | — | — | — | |
| MoDEBackbone=Qwen2.5-7B, Fusion=none2026.01 | 83.7 | — | 6.78 | — | — | — | |
| COMPACTStudent Architecture=Qwen2.5-7B, Distillation Method=COMPACT2026.01 | 83.59 | — | — | — | — | — | |
| Qwen2.5-7B-InstructStudent Architecture=Qwen2.5-7B, Distillation Method=Zero-shot2026.01 | 83.45 | — | — | — | — | — | |
| BaseBackbone=Qwen2.5-7B-Instruct, Fusion=none2026.01 | 83.45 | — | 4.46 | — | — | — | |
| SBS-KDStudent Architecture=Qwen2.5-7B, Distillation Method=SBS-KD2026.01 | 83.2 | — | — | — | — | — | |
| MoTStudent Architecture=Qwen2.5-7B, Distillation Method=MoT2026.01 | 83.2 | — | — | — | — | — | |
| SbSBackbone=Qwen2.5-7B, Fusion=none2026.01 | 83.2 | — | 5.97 | — | — | — | |
| SAC Single-taskModel=Qwen2-72B (Instruct)2024.11 | 82.72 | — | — | — | — | — | |
| SAC Multi-taskModel=Qwen2-72B (Instruct)2024.11 | 82.56 | — | — | — | — | — | |
| No ControlModel=Qwen2-72B (Instruct)2024.11 | 82.39 | — | — | — | — | — | |
| Thought-ICS-SModel=OSS-120B2026.02 | 82 | — | — | — | — | — | |
| Thought-ICS-AModel=OSS-120B2026.02 | 82 | — | — | — | — | — | |
| Token-ICSModel=OSS-20B2026.02 | 82 | — | — | — | — | — | |
| Qwen3-4BParams=4B2025.12 | 81.9 | — | — | — | — | — | |
| MCC-KDStudent Architecture=Qwen2.5-7B, Distillation Method=MCC-KD2026.01 | 81.72 | — | — | — | — | — | |
| MCCBackbone=Qwen2.5-7B, Fusion=none2026.01 | 81.72 | — | 6.76 | — | — | — | |
| MINDBackbone=Qwen2.5-7B, Fusion=without2026.01 | 81.05 | — | 9.3 | — | — | — | |
| Thought-ICS-SModel=OSS-20B2026.02 | 81 | — | — | — | — | — | |
| Thought-ICS-AModel=OSS-20B2026.02 | 81 | — | — | — | — | — | |
| 5-shot CoTModel=PaLM-540B, Setting=5-shot, # Paths=142023.05 | 80.7 | — | — | — | — | — | |
| 8-shot CoTModel=PaLM-540B, Setting=Prev, # Paths=402023.05 | 80.7 | — | — | — | — | — | |
| CommitteeStudent Architecture=Qwen2.5-7B, Distillation Method=Committee2026.01 | 80.52 | — | — | — | — | — | |
| COLLATEBackbone=open-llama-v2-7b2025.06 | 80.13 | — | — | — | — | — | |
| Self-RefineModel=OSS-120B2026.02 | 80 | — | — | — | — | — | |
| MV@10Model=OSS-20B2026.02 | 80 | — | — | — | — | — | |
| Self-reMAS Topology=complete2025.11 | 79.73 | — | — | — | — | — | |
| Auto-CoTModel=PaLM-540B, Setting=0-shot, # Paths=142023.05 | 79.4 | — | — | — | — | — | |
| MV@10Model=LLaMA-8B2026.02 | 79 | — | — | — | — | — | |
| SAC Single-taskModel=Qwen2-7B (Instruct)2024.11 | 78.95 | — | — | — | — | — | |
| SPINCategory=Training-driven Rationale Enhancement, Backbone=open-llama-v2-7b2025.06 | 78.76 | — | — | — | — | — | |
| Self-reMAS Topology=Star2025.11 | 78.65 | — | — | — | — | — | |
| No ControlModel=Qwen2-7B (Instruct)2024.11 | 78.62 | — | — | — | — | — | |
| Self-RefineModel=OSS-20B2026.02 | 78 | — | — | — | — | — | |
| CoVeModel=LLaMA-70B2026.02 | 78 | — | — | — | — | — | |
| Self-reMAS Topology=chain2025.11 | 77.84 | — | — | — | — | — | |
| Self-AgreementBackbone=GPT-3.5-turbo, Scenario=Question type and answer format unknown2023.11 | 77.8 | — | — | — | — | — | |
| CoTModel=OSS-20B2026.02 | 77 | — | — | — | — | — | |
| InitModel=OSS-20B2026.02 | 77 | — | — | — | — | — | |
| Full2025.12 | 77 | — | — | — | — | — | |
| H2OKV cache budget=3842025.12 | 77 | — | — | — | — | — | |
| H2OKV cache budget=5122025.12 | 77 | — | — | — | — | — | |
| KnormKV cache budget=2562025.12 | 77 | — | — | — | — | — | |
| RKVKV cache budget=5122025.12 | 77 | — | — | — | — | — | |
| StreamingLLMKV cache budget=5122025.12 | 77 | — | — | — | — | — | |
| SAC Multi-taskModel=Qwen2-7B (Instruct)2024.11 | 76.9 | — | — | — | — | — | |
| Self-reMAS Topology=tree2025.11 | 76.62 | — | — | — | — | — | |
| Llama2-70BModel Type=Single Model, Output Length=1002024.05 | 76.58 | — | — | 43,285.73 | — | — | |
| Few-shot Manual-CoTLLM=GPT-3.5-turbo, Evaluation Protocol=Few-shot2024.04 | 76.5 | — | — | — | — | — | |
| COSPModel=PaLM-540B, Setting=0-shot, # Paths=7+72023.05 | 76.4 | — | — | — | — | — |