List your company
ResearchFormerly WizWand

Reasoning on GSM8K (Accuracy)

98.8Accuracy (GSM8K)LoRA
4.6829.11553.5577.985Nov 17, 2025Dec 18, 2025Jan 19, 2026Feb 20, 2026Mar 23, 2026Apr 24, 2026May 26, 2026
Updated 5mo ago

Evaluation results

MethodLinks
LoRABackbone=Qwen2.5-7B-Instruct2026.0598.8
EVA (text)Backbone=Qwen2.5-7B-Instruct2026.0598.8
VanillaBackbone=Qwen2.5-7B-Instruct2026.0598.7
SafeDecodingBackbone=Qwen2.5-7B-Instruct2026.0598.7
LEDBackbone=Qwen2.5-7B-Instruct2026.0598.7
Circuit BreakersBackbone=Qwen2.5-7B-Instruct2026.0598.7
EVA (text)Backbone=Llama3.1-8B-Instruct2026.0598.5
VanillaBackbone=Llama3.1-8B-Instruct2026.0598.3
VanillaBackbone=Vicuna-7B-v1.52026.0598.2
LoRABackbone=Llama3.1-8B-Instruct2026.0598.2
EVA (text)Backbone=Vicuna-7B-v1.52026.0598.1
LEDBackbone=Llama3.1-8B-Instruct2026.0598.1
VanillaBackbone=Llama2-7B-chat2026.0597.7
Circuit BreakersBackbone=Llama3.1-8B-Instruct2026.0597.7
LoRABackbone=Vicuna-7B-v1.52026.0597.6
LoRABackbone=Llama2-7B-chat2026.0597.6
SafeDecodingBackbone=Llama2-7B-chat2026.0597.6
LEDBackbone=Vicuna-7B-v1.52026.0597.4
EVA (text)Backbone=Llama2-7B-chat2026.0597.4
LEDBackbone=Llama2-7B-chat2026.0597.3
Circuit BreakersBackbone=Vicuna-7B-v1.52026.0597.2
Circuit BreakersBackbone=Llama2-7B-chat2026.0597.1
SafeDecodingBackbone=Vicuna-7B-v1.52026.0596.9
SafeDecodingBackbone=Llama3.1-8B-Instruct2026.0596.9
DALAToken Cons.=6.20 × 10⁶2025.1196.2
AgentPrune-RToken Cons.=7.50 × 10⁶2025.1195.8
EVA (text)Backbone=Mistral-7B-Instruct-v0.22026.0595.5
LoRABackbone=Mistral-7B-Instruct-v0.22026.0594.6
Circuit BreakersBackbone=Mistral-7B-Instruct-v0.22026.0594.3
VanillaBackbone=Mistral-7B-Instruct-v0.22026.0594.1
LEDBackbone=Mistral-7B-Instruct-v0.22026.0593.7
PHPToken Cons.=2.60 × 10⁷2025.1192.5
LLM-DebateToken Cons.=2.20 × 10⁷2025.1190.2
DyLANToken Cons.=1.40 × 10⁷2025.1188.2
DeepSeek-R1-Distill-Qwen-14B (Reasoning)Model Family=Qwen2.5-14B2026.0186.43
SafeDecodingBackbone=Mistral-7B-Instruct-v0.22026.0586.3
ReasonAnyModel Family=Qwen2.5-14B2026.0185.44
Rubric-grounded GRPOCheckpoint=Best by held-out rubric reward, Backbone=Llama-3.1-8B-Instruct2026.0585.44
VanillaToken Cons.=3.50 × 10⁶2025.1185.4
Llama-3.1-8B-InstructCheckpoint=Base2026.0584.53
Task ArithmeticModel Family=Qwen2.5-14B2026.0181.96
TIESModel Family=Qwen2.5-14B2026.0179.76
LEDModel Family=Qwen2.5-14B2026.0176.42
Qwen2.5-14B-Instruct (Safety)Model Family=Qwen2.5-14B2026.0175.74
DAREModel Family=Qwen2.5-14B2026.0164.9
Teacher (Qwen3-0.6B)Role=Teacher, Model=Qwen3-0.6B, Evaluation Protocol=generation-based evaluation2026.0357.5
MobileMoE-LActive Parameters=922M, Total Parameters=5.3B, Few-shot count=8-shot2026.0555.7
FuseLLMModel Family=Qwen2.5-14B2026.0152.46
MobileMoE-MActive Parameters=528M, Total Parameters=2.8B, Few-shot count=8-shot2026.0551.6
LinearModel Family=Qwen2.5-14B2026.0150.57
MobileMoE-SActive Parameters=272M, Total Parameters=1.3B, Few-shot count=8-shot2026.0536.2
Hybrid KDASequence Mixer=KDA, Attention Layers=7, Evaluation Protocol=generation-based evaluation2026.0334.5
Hybrid MambaSequence Mixer=Mamba2, Attention Layers=7, Evaluation Protocol=generation-based evaluation2026.0328.3
Pure KDASequence Mixer=KDA, Attention Layers=0, Evaluation Protocol=generation-based evaluation2026.0322.8
Pure MambaSequence Mixer=Mamba2, Attention Layers=0, Evaluation Protocol=generation-based evaluation2026.038.3