List your company

Commonsense Reasoning on HellaSwag (Accuracy, Delta, and Confidence Intervals)

89.3AccuracyMoE-Sieve (Qwen1.5-MoE-A2.7B)
14.4233.8653.372.74Mar 25, 2026Mar 28, 2026Apr 1, 2026Apr 5, 2026Apr 9, 2026Apr 13, 2026Apr 17, 2026

Evaluation results

MethodLinks
MoE-Sieve (Qwen1.5-MoE-A2.7B)Condition=Hot (25%)2026.0389.30.730.53—
Qwen1.5-MoE-A2.7BCondition=Full LoRA2026.0388.5———
MoE-Sieve (OLMoE-1B-7B)Condition=Hot (25%)2026.0380.70.170.71—
OLMoE-1B-7BCondition=Full LoRA2026.0380.5———
NAG_Qwen3-1.7BTarget Setting=Single-Target, Backbone=Qwen3-1.7B2026.0460.6———
NAG_SmolLM3-3BTarget Setting=Single-Target, Backbone=SmolLM3-3B2026.0459.8———
NAG_Llama-3.2-3BTarget Setting=Single-Target, Backbone=Llama-3.2-3B2026.0458.6———
NAG_Qwen3-1.7BTarget Setting=Multi-Target, Backbone=Qwen3-1.7B2026.0457.8———
BETRTarget Setting=Single-Target2026.0457.5———
FineWeb-Edu2026.0455.3———
NAG_SmolLM3-3BTarget Setting=Multi-Target, Backbone=SmolLM3-3B2026.0455.2———
NAG_Llama-3.2-3BTarget Setting=Multi-Target, Backbone=Llama-3.2-3B2026.0454.9———
Random2026.0451.6———
BETRTarget Setting=Multi-Target2026.0449.3———
Teacher (Qwen3-0.6B)Role=Teacher, Model=Qwen3-0.6B, Evaluation Protocol=generation-based evaluation2026.0339———
Hybrid KDASequence Mixer=KDA, Attention Layers=7, Evaluation Protocol=generation-based evaluation2026.0330———
Hybrid MambaSequence Mixer=Mamba2, Attention Layers=7, Evaluation Protocol=generation-based evaluation2026.0329———
Pure KDASequence Mixer=KDA, Attention Layers=0, Evaluation Protocol=generation-based evaluation2026.0320.3———
Pure MambaSequence Mixer=Mamba2, Attention Layers=0, Evaluation Protocol=generation-based evaluation2026.0317.3———
FLOATDOORScenario=S1, Backbone=Qwen3-4B2026.06—0.05——
FLOATDOORScenario=S2, Backbone=Qwen3-4B2026.06—0.01——
FLOATDOORScenario=S3, Backbone=Qwen3-4B2026.06—0.11——
FLOATDOORScenario=M1, Backbone=Qwen3-4B2026.06—0.2——
FLOATDOORScenario=M2, Backbone=Qwen3-4B2026.06—0.38——