List your company
ResearchFormerly WizWand

Multi-task Knowledge Understanding on MMLU-Pro

59.64Mean AccuracyPolicyAlign
-1.345614.487230.3246.1528Feb 16, 2026Mar 9, 2026Mar 30, 2026Apr 21, 2026May 12, 2026Jun 2, 2026Jun 24, 2026
Updated 4mo ago

Evaluation results

MethodLinks
PolicyAlignBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Policy-Based Safety Alignment2026.0659.64
BaseBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Base2026.0659.28
NSPOBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Null-space constrained policy optimization2026.0658.74
ICLBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=In-Context Learning (System Prompt)2026.0658.52
GRPO+PolicyBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Group Relative Policy Optimization with policy-conditioned reward2026.0658.44
AlphaAlignBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=RL-based alignment2026.0658.07
SFTBackbone=Qwen2.5-14B-Instruct, Alignment Strategy=Supervised Fine-Tuning2026.0653.24
BaseBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Base2026.0651.7
ICLBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=In-Context Learning (System Prompt)2026.0651.61
PolicyAlignBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Policy-Based Safety Alignment2026.0651.25
NSPOBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Null-space constrained policy optimization2026.0650.86
GRPO+PolicyBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Group Relative Policy Optimization with policy-conditioned reward2026.0650.69
AlphaAlignBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=RL-based alignment2026.0648.37
SFTBackbone=Qwen2.5-7B-Instruct, Alignment Strategy=Supervised Fine-Tuning2026.0645.95
OPDTraining Budget=sufficient, Rollout=full, Zero-shot=true2026.0241.3
OPD (prefix scheduling)Training Budget=sufficient, Zero-shot=true2026.0239.4
AceMADBase Model=Qwen3-235B-A22B-Instruct, Iteration Rounds (T)=5, Number of Agents (N)=52026.0336.54
Single AgentBase Model=Qwen3-235B-A22B-Instruct, Number of Agents (N)=52026.0335.93
ICLBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=In-Context Learning (System Prompt)2026.0632.21
PolicyAlignBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Policy-Based Safety Alignment2026.0631.43
GRPO+PolicyBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Group Relative Policy Optimization with policy-conditioned reward2026.0631.27
BaseBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Base2026.0631.07
NSPOBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Null-space constrained policy optimization2026.0630.87
OPDSteps=10, Training Budget=limited, Zero-shot=true2026.0230.7
AlphaAlignBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=RL-based alignment2026.0629.88
SeqKDTraining Budget=sufficient, Zero-shot=true2026.0229.6
Decentralized MADBase Model=Qwen3-235B-A22B-Instruct, Number of Agents (N)=52026.0328.9
SeqKDSteps=300, Training Budget=limited, Zero-shot=true2026.0228.5
SFTBackbone=LLaMA-3.2-3B-Instruct, Alignment Strategy=Supervised Fine-Tuning2026.0628.22
AceMADBase Model=Qwen3-235B-A22B-Instruct, Iteration Rounds (T)=3, Number of Agents (N)=52026.0327.06
Centralized MADBase Model=Qwen3-235B-A22B-Instruct, Number of Agents (N)=52026.0324.01
AceMADBase Model=Qwen3-235B-A22B-Instruct, Iteration Rounds (T)=2, Number of Agents (N)=52026.0314.37
Majority VotingBase Model=Qwen3-235B-A22B-Instruct, Number of Agents (N)=52026.0314.07
Sparse MADBase Model=Qwen3-235B-A22B-Instruct, Number of Agents (N)=52026.0313.91
OPDPrefix Length=2048, Training Budget=limited, Zero-shot=true2026.0212
Qwen3-1.7B-BaseZero-shot=true2026.029.1
Centralized MADBase Model=GPT-4o-mini, Number of Agents (N)=52026.038.72
Sparse MADBase Model=GPT-4o-mini, Number of Agents (N)=52026.038.72
AceMADBase Model=GPT-4o-mini, Iteration Rounds (T)=3, Number of Agents (N)=52026.038.72
AceMADBase Model=GPT-4o-mini, Iteration Rounds (T)=2, Number of Agents (N)=52026.038.1
AceMADBase Model=GPT-4o-mini, Iteration Rounds (T)=5, Number of Agents (N)=52026.037.95
Decentralized MADBase Model=GPT-4o-mini, Number of Agents (N)=52026.036.88
Single AgentBase Model=GPT-4o-mini, Number of Agents (N)=52026.036.42
Majority VotingBase Model=GPT-4o-mini, Number of Agents (N)=52026.035.5
OPDPrefix Length=1024, Training Budget=limited, Zero-shot=true2026.021.9
OPDPrefix Length=256, Training Budget=limited, Zero-shot=true2026.021.5
OPDPrefix Length=512, Training Budget=limited, Zero-shot=true2026.021