List your company
ResearchFormerly WizWand

Adversarial Language Modeling on WikiText-2

5.44Perplexity (Baseline)A3
4.8169.02813.2417.452May 19, 2025Jun 20, 2025Jul 22, 2025Aug 23, 2025Sep 24, 2025Oct 26, 2025Nov 27, 2025
Updated 5mo ago

Evaluation results

MethodLinks
A3Compression=10%, Model=phi-3-medium-4k-instruct (14B)2025.055.44—1.14
A3Compression=20%, Model=phi-3-medium-4k-instruct (14B)2025.056.4—2.1
SVD-LLMCompression=10%, Model=phi-3-medium-4k-instruct (14B)2025.056.81—2.5
SVD-LLMCompression=20%, Model=phi-3-medium-4k-instruct (14B)2025.058.14—3.83
GDF-BFABackbone=Llama2-13B, # of bit flips=3, Quantization=INT82025.1111.02544.99—
Llama2-13BModel=Llama2-13B, # of bit flips=3, Quantization precision=4-bit, Attack=GDF-BFA2025.1111.44105.24—
GDF-BFABackbone=Mistral-7B-v0.3, # of bit flips=1, Quantization=INT82025.1112.181,752.41—
GDF-BFABackbone=Llama2-7B, # of bit flips=1, Quantization=INT82025.1112.321,493.07—
Llama2-7BModel=Llama2-7B, # of bit flips=1, Quantization precision=4-bit, Attack=GDF-BFA2025.1113.88,103.08—
GDF-BFABackbone=Llama3-8B, # of bit flips=1, Quantization=INT82025.1114.4632,048.31—
Llama3-8BModel=Llama3-8B, # of bit flips=1, Quantization precision=4-bit, Attack=GDF-BFA2025.1116.65344,551.89—
GDF-BFABackbone=Qwen3-8B, # of bit flips=6, Quantization=INT82025.1121.04391.01—