List your company

Mathematical Reasoning on GSM-8K (Accuracy)

97.3AccuracyFACT-E(Standard)
38.061653.440868.8284.1992May 26, 2025Jul 18, 2025Sep 10, 2025Nov 3, 2025Dec 27, 2025Feb 19, 2026Apr 14, 2026

Evaluation results

MethodLinks
FACT-E(Standard)Backbone=DeepSeek-V32026.0497.3
FACT-E(Lightweight)Backbone=DeepSeek-V32026.0497.2
CONSISTENCYBackbone=DeepSeek-V32026.0496.4
CoTBackbone=DeepSeek-V32026.0496
REFLECTBackbone=DeepSeek-V32026.0495.8
POLISHBackbone=DeepSeek-V32026.0495.6
DENOISEBackbone=DeepSeek-V32026.0495.4
FACT-EModel=DeepSeek-V3, K=52026.0495
FACT-EModel=Qwen3, K=52026.0494.8
BASEModel=DeepSeek-V3, K=52026.0494.6
CONSIST.Model=DeepSeek-V3, K=52026.0494.4
PaTaRMPolicy Model=Qwen3-8B2025.1094.3
FACT-E(Lightweight)Backbone=Gpt-4o-mini2026.0493.98
DENOISEModel=Qwen3, K=52026.0493.8
REFLECTModel=DeepSeek-V3, K=52026.0493.8
Skywork-Llama-3.1-8BPolicy Model=Qwen3-8B2025.1093.7
DENOISEModel=DeepSeek-V3, K=52026.0493.6
POLISHModel=Qwen3, K=52026.0493.6
FACT-E(Standard)Backbone=Qwen3-14B2026.0493.4
CoTBackbone=Qwen3-14B2026.0493.2
DENOISEBackbone=Qwen3-14B2026.0493.2
CONSISTENCYBackbone=Qwen3-14B2026.0493.2
FACT-E(Lightweight)Backbone=Qwen3-14B2026.0493.2
BASEModel=4o-mini, K=52026.0493.2
BASEModel=Qwen3, K=52026.0493.2
DENOISEModel=4o-mini, K=52026.0493
CONSISTENCYBackbone=Gpt-4o-mini2026.0492.8
REFLECTModel=Qwen3, K=52026.0492.8
POLISHBackbone=Gpt-4o-mini2026.0492.6
REFLECTBackbone=Qwen3-14B2026.0492.6
CoTBackbone=Gpt-4o-mini2026.0492.4
POLISHBackbone=Qwen3-14B2026.0492.4
CONSIST.Model=Qwen3, K=52026.0492.4
DENOISEBackbone=Gpt-4o-mini2026.0492.2
FACT-EModel=4o-mini, K=52026.0492.2
REFLECTBackbone=Gpt-4o-mini2026.0492.1
CONSIST.Model=4o-mini, K=52026.0492
FACT-E(Standard)Backbone=Gpt-4o-mini2026.0491.8
POLISHModel=4o-mini, K=52026.0491.6
POLISHModel=DeepSeek-V3, K=52026.0491.6
REFLECTModel=4o-mini, K=52026.0491
Rule-based RewardPolicy Model=Qwen3-8B2025.1090.6
Base ModelPolicy Model=Qwen3-8B2025.1089.8
LLAMA-4-SCOUTPARAMS=109B, Variant=FCSD2025.1189.8
LLAMA-4-SCOUTPARAMS=109B, Variant=ORIGINAL2025.1188.6
QWEN-3-30B MOEPARAMS=30B, Variant=FCSD2025.1187.6
QWEN-3-30B MOEPARAMS=30B, Variant=ORIGINAL2025.1186.9
FACT-E(Lightweight)Backbone=ChatGPT (Gpt-3.5-turbo)2026.0481.4
PaTaRMPolicy Model=Qwen3-0.6B2025.1081
CONSISTENCYBackbone=ChatGPT (Gpt-3.5-turbo)2026.0480.6
LLAMA-4-SCOUTPARAMS=109B, Variant=SFT2025.1178.6
REFLECTBackbone=ChatGPT (Gpt-3.5-turbo)2026.0478.4
Rule-based RewardPolicy Model=Qwen3-0.6B2025.1078.4
PiCaBackbone=Gemma-7B, #Params=10.22M, r=2562025.0578.39
Full-FTBackbone=Gemma-7B, #Params=8.5B2025.0578.09
CoTBackbone=ChatGPT (Gpt-3.5-turbo)2026.0477.6
FACT-EModel=ChatGpt, K=52026.0477.4
Skywork-Llama-3.1-8BPolicy Model=Qwen3-0.6B2025.1077.3
POLISHBackbone=ChatGPT (Gpt-3.5-turbo)2026.0477
FACT-E(Standard)Backbone=ChatGPT (Gpt-3.5-turbo)2026.0477
DENOISEModel=ChatGpt, K=52026.0477
Base ModelPolicy Model=Qwen3-0.6B2025.1077
SVFT-RBackbone=Gemma-7B, #Params=19.8M, d=162025.0576.81
DENOISEBackbone=ChatGPT (Gpt-3.5-turbo)2026.0476.8
LoRABackbone=Gemma-7B, #Params=68.8M, r=322025.0576.57
Full-FTBackbone=LLaMA-3-8B, #Params=8.0B2025.0576.57
REFLECTModel=ChatGpt, K=52026.0476.2
PiCaBackbone=LLaMA-3-8B, #Params=11.01M, r=2562025.0576.12
QWEN-3-30B MOEPARAMS=30B, Variant=SFT2025.1176
SVFT-RBackbone=LLaMA-3-8B, #Params=13.1M, d=122025.0575.9
LoRABackbone=LLaMA-3-8B, #Params=56.6M, r=322025.0575.89
DoRABackbone=LLaMA-3-8B, #Params=29.1M, r=162025.0575.66
DoRABackbone=Gemma-7B, #Params=35.5M, r=162025.0574.52
DoRABackbone=Gemma-7B, #Params=3.26M, r=12025.0574.37
PiCaBackbone=Gemma-7B, #Params=0.64M, r=322025.0574.3
BASEModel=ChatGpt, K=52026.0474.2
PiCaBackbone=LLaMA-3-8B, #Params=1.38M, r=322025.0573.54
SVFT-PBackbone=Gemma-7B, #Params=0.43M2025.0573.5
LoRABackbone=Gemma-7B, #Params=0.82M, r=12025.0572.4
POLISHModel=ChatGpt, K=52026.0471.8
BOFTBackbone=Gemma-7B, #Params=2.90M, config=b=8, m=22025.0571.79
CONSIST.Model=ChatGpt, K=52026.0471.2
VeRABackbone=Gemma-7B, #Params=0.43M, r=10242025.0571.11
DEEPSEEK-V2-LITEPARAMS=16B, Variant=ORIGINAL2025.1170.1
SVFT-PBackbone=LLaMA-3-8B, #Params=0.48M2025.0569.22
LoRABackbone=LLaMA-3-8B, #Params=1.77M, r=12025.0568.84
DoRABackbone=LLaMA-3-8B, #Params=2.55M, r=12025.0568.3
BOFTBackbone=LLaMA-3-8B, #Params=4.35M, config=b=8, m=22025.0567.09
DEEPSEEK-V2-LITEPARAMS=16B, Variant=FCSD2025.1164.2
VeRABackbone=LLaMA-3-8B, #Params=0.98M, r=10242025.0563.76
DEEPSEEK-V2-LITEPARAMS=16B, Variant=SFT2025.1154.3
PiCaBackbone=Gemma-2B, #Params=5.37M, r=2562025.0552.77
Full-FTBackbone=Gemma-2B, #Params=2.5B2025.0552.69
SVFT-RBackbone=Gemma-2B, #Params=6.35M, d=162025.0550.03
DoRABackbone=Gemma-2B, #Params=13.5M, r=162025.0544.27
LoRABackbone=Gemma-2B, #Params=26.2M, r=322025.0543.06
DeepSeek V2-LiteAttention=MLA2026.0441.47
PiCaBackbone=Gemma-2B, #Params=0.67M, r=322025.0541.32
LCA2026.0441.17
SVFT-PBackbone=Gemma-2B, #Params=0.19M2025.0540.34