List your company
ResearchFormerly WizWand

Accuracy on ARC Challenge (Reasoning)

0.946Accuracy (ARC)GRPO
0.284560.456280.6280.79972Sep 8, 2025Oct 25, 2025Dec 12, 2025Jan 29, 2026Mar 18, 2026May 5, 2026Jun 22, 2026
Updated 4mo ago

Evaluation results

MethodLinks
GRPOBase Model=Qwen3-4B2025.100.946
SCOPE-RLBase Model=Qwen3-4B2025.100.943
Entropy-RegBase Model=Qwen3-4B2025.100.939
Entropy-AdvBase Model=Qwen3-4B2025.100.936
KL-covBase Model=Qwen3-4B2025.100.933
BaseBase Model=Qwen3-4B2025.100.929
DAPOBase Model=Qwen3-4B2025.100.926
CISPOBase Model=Qwen3-4B2025.100.923
QwenCheckpoint=SFTMaxOOD2025.090.92
SCOPE-RLBase Model=Qwen2.5-7B2025.100.916
DAPOBase Model=Qwen2.5-7B2025.100.909
KL-covBase Model=Qwen2.5-7B2025.100.909
Entropy-AdvBase Model=Qwen2.5-7B2025.100.906
CISPOBase Model=Qwen2.5-7B2025.100.899
Entropy-RegBase Model=Qwen2.5-7B2025.100.896
GRPOBase Model=Qwen2.5-7B2025.100.893
QwenCheckpoint=SFTEnd2025.090.88
QwenCheckpoint=RLEnd2025.090.88
BaseBase Model=Qwen2.5-7B2025.100.863
DAPOBase Model=Qwen2.5-Math-7B2025.100.809
CISPOBase Model=Qwen2.5-Math-7B2025.100.809
SCOPE-RLBase Model=Qwen2.5-Math-7B2025.100.809
Entropy-AdvBase Model=Qwen2.5-Math-7B2025.100.803
LLaMACheckpoint=SFTMaxOOD2025.090.8
LLaMACheckpoint=RLEnd2025.090.8
KL-covBase Model=Qwen2.5-Math-7B2025.100.799
Entropy-RegBase Model=Qwen2.5-Math-7B2025.100.796
GRPOBase Model=Qwen2.5-Math-7B2025.100.789
LLaMACheckpoint=SFTEnd2025.090.7
BaseBase Model=Qwen2.5-Math-7B2025.100.699
BF16-BaselineEff. Bits=16, Model=Qwen3-8B2026.060.5657
GRINQH-6b-RTNEff. Bits=3.05, Model=Qwen3-8B2026.060.5614
BF16-BaselineEff. Bits=16, Model=Llama-3.1-8B-Instruct2026.060.5589
GRINQH-8b-RTNEff. Bits=3.96, Model=Llama-3.1-8B-Instruct2026.060.558
GRINQH-4b-GPTQEff. Bits=3.02, Model=Qwen3-8B2026.060.552
GRINQH-8b-RTNEff. Bits=3.06, Model=Qwen3-8B2026.060.5512
GRINQH-8b-RTNEff. Bits=2.95, Model=Llama-3.1-8B-Instruct2026.060.5503
GRINQH-4b-GPTQEff. Bits=3.01, Model=Llama-3.1-8B-Instruct2026.060.5427
MatGPTQ-EP-Mix’n’MatchEff. Bits=4.00, Model=Llama-3.1-8B-Instruct2026.060.5384
GRINQH-6b-RTNEff. Bits=2.98, Model=Llama-3.1-8B-Instruct2026.060.5358
GRINQH-8b-RTNEff. Bits=2.24, Model=Qwen3-8B2026.060.5299
GRINQH-6b-RTNEff. Bits=3.881, Model=Llama-3.1-8B-Instruct2026.060.529
MatGPTQEff. Bits=4.00, Model=Llama-3.1-8B-Instruct2026.060.5282
GRINQH-4b-GPTQEff. Bits=2.09, Model=Qwen3-8B2026.060.5171
MatGPTQEff. Bits=3.00, Model=Qwen3-8B2026.060.4872
MatGPTQ-EP-Mix’n’MatchEff. Bits=3.00, Model=Llama-3.1-8B-Instruct2026.060.4753
MatGPTQ-EP-Mix’n’MatchEff. Bits=3.00, Model=Qwen3-8B2026.060.4753
YOCO (CLSA)Model Scale=4B, Attention=Cross-Layer Sparse2026.060.465
MatGPTQEff. Bits=3.00, Model=Llama-3.1-8B-Instruct2026.060.4625
YOCO (Dense)Model Scale=4B2026.060.461
GRINQH-8b-RTNEff. Bits=2.08, Model=Llama-3.1-8B-Instruct2026.060.4608
TransformerModel Scale=4B2026.060.453
BaselineCompression Ratio=0.02026.050.38
SAES-SVD + PARSECompression Ratio=0.22026.050.37
SAES-SVD + PARSECompression Ratio=0.42026.050.31
SAES-SVD + PARSECompression Ratio=0.62026.050.31