Language Modeling on Perplexity
3.43Perplexity (PPL)SERQ (GPTQ)
Evaluation Results
| Method | Links | |
|---|---|---|
| SERQ (GPTQ)Backbone=LLaMA-2 70B, #Bits=W4A8, #Eff. (w)=3.432026.03 | 3.43 | |
| SERQ (RTN)Backbone=LLaMA-2 70B, #Bits=W4A8, #Eff. (w)=3.442026.03 | 3.44 | |
| baselineBackbone=LLaMA-2 70B, #Bits=FP16, #Eff. (w)=162026.03 | 3.53 | |
| L2QERBackbone=LLaMA-2 70B, #Bits=W4A8, #Eff. (w)=3.552026.03 | 3.55 | |
| SERQ (GPTQ)Backbone=LLaMA-2 70B, #Bits=W4A4, #Eff. (w)=3.642026.03 | 3.64 | |
| SERQ (RTN)Backbone=LLaMA-2 70B, #Bits=W4A4, #Eff. (w)=3.652026.03 | 3.65 | |
| SERQ-MXFP4Backbone=LLaMA-2 70B, #Bits=W4A4, #Eff. (w)=3.792026.03 | 3.79 | |
| L2QER-MXFP4Backbone=LLaMA-2 70B, #Bits=W4A4, #Eff. (w)=3.812026.03 | 3.81 | |
| L2QERBackbone=LLaMA-2 70B, #Bits=W4A4, #Eff. (w)=4.552026.03 | 4.55 | |
| Clean FT (No WM)Model=Llama-2-7B, Condition=Original (Pre-Attack)2026.03 | 5.3 | |
| baselineBackbone=LLaMA-2 7B, #Bits=FP16, #Eff. (w)=162026.03 | 5.47 | |
| Pre-trainedModel Backbone=LLaMA2-7B, Precision=FP162025.11 | 5.47 | |
| SERQ (GPTQ)Backbone=LLaMA-2 7B, #Bits=W4A8, #Eff. (w)=4.242026.03 | 5.59 | |
| SERQ (RTN)Backbone=LLaMA-2 7B, #Bits=W4A8, #Eff. (w)=4.242026.03 | 5.64 | |
| EmMarkModel=Llama-2-7B, Condition=Attacked (Post-Attack)2026.03 | 5.73 | |
| Weight Quant.Model=Llama-2-7B, Condition=Attacked (Post-Attack)2026.03 | 5.78 | |
| L2QERBackbone=LLaMA-2 7B, #Bits=W4A8, #Eff. (w)=4.352026.03 | 5.83 | |
| Clean FT (No WM)Model=Llama-3-8B, Condition=Original (Pre-Attack)2026.03 | 5.83 | |
| Full DIFSWModel=Llama-2-7B, Condition=Original (Pre-Attack)2026.03 | 5.91 | |
| Full DIFSWModel=Llama-2-7B, Condition=Attacked (Post-Attack)2026.03 | 5.91 | |
| SERQ (GPTQ)Backbone=LLaMA-2 7B, #Bits=W4A4, #Eff. (w)=4.242026.03 | 5.97 | |
| SERQ (RTN)Backbone=LLaMA-2 7B, #Bits=W4A4, #Eff. (w)=4.242026.03 | 6.03 | |
| Weight Quant.Model=Llama-3-8B, Condition=Attacked (Post-Attack)2026.03 | 6.03 | |
| EmMarkModel=Llama-3-8B, Condition=Attacked (Post-Attack)2026.03 | 6.09 | |
| Full DIFSWModel=Llama-3-8B, Condition=Attacked (Post-Attack)2026.03 | 6.12 | |
| baselineBackbone=LLaMA-3 8B, #Bits=FP16, #Eff. (w)=162026.03 | 6.13 | |
| Pre-trainedModel Backbone=LLaMA3.1-8B, Precision=FP162025.11 | 6.14 | |
| Full DIFSWModel=Llama-3-8B, Condition=Original (Pre-Attack)2026.03 | 6.2 | |
| SERQ-MXFP4Backbone=LLaMA-2 7B, #Bits=W4A4, #Eff. (w)=4.372026.03 | 6.22 | |
| Weight Quant.Model=Mistral-7B, Condition=Attacked (Post-Attack)2026.03 | 6.23 | |
| EmMarkModel=Mistral-7B, Condition=Attacked (Post-Attack)2026.03 | 6.24 | |
| L2QER-MXFP4Backbone=LLaMA-2 7B, #Bits=W4A4, #Eff. (w)=4.372026.03 | 6.3 | |
| Clean FT (No WM)Model=Qwen2.5-7B, Condition=Original (Pre-Attack)2026.03 | 6.35 | |
| SERQ (GPTQ)Backbone=LLaMA-3 8B, #Bits=W4A8, #Eff. (w)=6.522026.03 | 6.52 | |
| Full DIFSWModel=Mistral-7B, Condition=Attacked (Post-Attack)2026.03 | 6.59 | |
| Naive Top-kModel=Llama-3-8B, Condition=Attacked (Post-Attack)2026.03 | 6.64 | |
| Weight Quant.Model=Qwen2.5-7B, Condition=Attacked (Post-Attack)2026.03 | 6.66 | |
| EmMarkModel=Qwen2.5-7B, Condition=Attacked (Post-Attack)2026.03 | 6.68 | |
| SERQ (RTN)Backbone=LLaMA-3 8B, #Bits=W4A8, #Eff. (w)=6.712026.03 | 6.71 | |
| Naive Top-kModel=Llama-2-7B, Condition=Attacked (Post-Attack)2026.03 | 6.75 | |
| Naive Top-kModel=Mistral-7B, Condition=Attacked (Post-Attack)2026.03 | 6.93 | |
| Fine-tunedModel Backbone=LLaMA2-7B, Precision=FP162025.11 | 6.94 | |
| Pre-trainedModel Backbone=Qwen2-7B, Precision=FP162025.11 | 7.13 | |
| L2QERBackbone=LLaMA-3 8B, #Bits=W4A8, #Eff. (w)=7.162026.03 | 7.16 | |
| Fine-tunedModel Backbone=LLaMA3.1-8B, Precision=FP162025.11 | 7.23 | |
| Full DIFSWModel=Qwen2.5-7B, Condition=Original (Pre-Attack)2026.03 | 7.24 | |
| LLM.int4()Backbone=LLaMA-2 7B, #Bits=W4A8, #Eff. (w)=-†2026.03 | 7.28 | |
| OSTQuantModel Backbone=LLaMA2-7B, Precision=W4A42025.11 | 7.28 | |
| Naive Top-kModel=Llama-2-7B, Condition=Original (Pre-Attack)2026.03 | 7.3 | |
| Naive Top-kModel=Llama-3-8B, Condition=Original (Pre-Attack)2026.03 | 7.35 | |
| L2QERBackbone=LLaMA-2 7B, #Bits=W4A4, #Eff. (w)=4.242026.03 | 7.37 | |
| CAQModel Backbone=LLaMA2-7B, Precision=W4A42025.11 | 7.56 | |
| Fine-tunedModel Backbone=Qwen2-7B, Precision=FP162025.11 | 7.6 | |
| SERQ-MXFP4Backbone=LLaMA-3 8B, #Bits=W4A4, #Eff. (w)=7.632026.03 | 7.63 | |
| Full DIFSWModel=Qwen2.5-7B, Condition=Attacked (Post-Attack)2026.03 | 7.66 | |
| QuaRotModel Backbone=LLaMA2-7B, Precision=W4A42025.11 | 7.67 | |
| SpinQuantModel Backbone=LLaMA2-7B, Precision=W4A42025.11 | 7.68 | |
| SERQ (GPTQ)Backbone=LLaMA-3 8B, #Bits=W4A4, #Eff. (w)=7.752026.03 | 7.75 | |
| Naive Top-kModel=Qwen2.5-7B, Condition=Attacked (Post-Attack)2026.03 | 7.79 | |
| baselineBackbone=LLaMA-3.2 3B, #Bits=FP16, #Eff. (w)=162026.03 | 7.81 | |
| L2QER-MXFP4Backbone=LLaMA-3 8B, #Bits=W4A4, #Eff. (w)=7.832026.03 | 7.83 | |
| DuQuantModel Backbone=LLaMA2-7B, Precision=W4A42025.11 | 7.88 | |
| Naive Top-kModel=Qwen2.5-7B, Condition=Original (Pre-Attack)2026.03 | 7.94 | |
| EmMarkModel=DeepSeek-7B, Condition=Attacked (Post-Attack)2026.03 | 7.95 | |
| baselineBackbone=Qwen-2.5-3B, #Bits=FP16, #Eff. (w)=162026.03 | 8.03 | |
| SERQ (RTN)Backbone=LLaMA-3 8B, #Bits=W4A4, #Eff. (w)=8.072026.03 | 8.07 | |
| SERQ (GPTQ)Backbone=LLaMA-3.2 3B, #Bits=W4A8, #Eff. (w)=4.242026.03 | 8.18 | |
| OSTQuantModel Backbone=Qwen2-7B, Precision=W4A42025.11 | 8.18 | |
| CAQModel Backbone=Qwen2-7B, Precision=W4A42025.11 | 8.23 | |
| Weight Quant.Model=DeepSeek-7B, Condition=Attacked (Post-Attack)2026.03 | 8.23 | |
| OSTQuantModel Backbone=LLaMA3.1-8B, Precision=W4A42025.11 | 8.29 | |
| SERQ (GPTQ)Backbone=Qwen-2.5-3B, #Bits=W4A8, #Eff. (w)=4.242026.03 | 8.36 | |
| SERQ (RTN)Backbone=LLaMA-3.2 3B, #Bits=W4A8, #Eff. (w)=4.242026.03 | 8.38 | |
| CAQModel Backbone=LLaMA3.1-8B, Precision=W4A42025.11 | 8.41 | |
| SpinQuantModel Backbone=Qwen2-7B, Precision=W4A42025.11 | 8.41 | |
| QuaRotModel Backbone=Qwen2-7B, Precision=W4A42025.11 | 8.43 | |
| SpinQuantModel Backbone=LLaMA3.1-8B, Precision=W4A42025.11 | 8.57 | |
| SERQ (RTN)Backbone=Qwen-2.5-3B, #Bits=W4A8, #Eff. (w)=4.242026.03 | 8.67 | |
| SERQ-MXFP4Backbone=LLaMA-3.2 3B, #Bits=W4A4, #Eff. (w)=4.372026.03 | 8.74 | |
| L2QERBackbone=LLaMA-3.2 3B, #Bits=W4A8, #Eff. (w)=4.352026.03 | 8.78 | |
| QuaRotModel Backbone=LLaMA3.1-8B, Precision=W4A42025.11 | 8.81 | |
| SERQ (GPTQ)Backbone=LLaMA-3.2 3B, #Bits=W4A4, #Eff. (w)=4.242026.03 | 9.15 | |
| DuQuantModel Backbone=LLaMA3.1-8B, Precision=W4A42025.11 | 9.22 | |
| SERQ (GPTQ)Backbone=Qwen-2.5-3B, #Bits=W4A4, #Eff. (w)=4.242026.03 | 9.35 | |
| SERQ (RTN)Backbone=LLaMA-3.2 3B, #Bits=W4A4, #Eff. (w)=4.242026.03 | 9.43 | |
| L2QER-MXFP4Backbone=LLaMA-3.2 3B, #Bits=W4A4, #Eff. (w)=4.372026.03 | 9.59 | |
| SERQ (RTN)Backbone=Qwen-2.5-3B, #Bits=W4A4, #Eff. (w)=4.242026.03 | 9.66 | |
| baselineBackbone=LLaMA-3.2 1B, #Bits=FP16, #Eff. (w)=162026.03 | 9.75 | |
| Weight Quant.Model=Llama-3-8B, Condition=Original (Pre-Attack)2026.03 | 9.77 | |
| SERQ-MXFP4Backbone=Qwen-2.5-3B, #Bits=W4A4, #Eff. (w)=4.372026.03 | 9.79 | |
| SERQ (GPTQ)Backbone=LLaMA-3.2 1B, #Bits=W4A8, #Eff. (w)=4.242026.03 | 10.45 | |
| L2QER-MXFP4Backbone=Qwen-2.5-3B, #Bits=W4A4, #Eff. (w)=4.372026.03 | 10.46 | |
| LLM.int4()Backbone=LLaMA-3 8B, #Bits=W4A8, #Eff. (w)=-†2026.03 | 10.82 | |
| SERQ (RTN)Backbone=LLaMA-3.2 1B, #Bits=W4A8, #Eff. (w)=4.242026.03 | 11.14 | |
| L2QERBackbone=LLaMA-3 8B, #Bits=W4A4, #Eff. (w)=11.442026.03 | 11.44 | |
| DensePrune%=0%, Evaluation=Zero-shot, Model=Mistral-7B2025.12 | 11.9 | |
| L2QERBackbone=LLaMA-3.2 1B, #Bits=W4A8, #Eff. (w)=4.352026.03 | 12.06 | |
| L2QERBackbone=Qwen-2.5-3B, #Bits=W4A8, #Eff. (w)=4.352026.03 | 12.25 | |
| LLM.int4()Backbone=LLaMA-3.2 3B, #Bits=W4A8, #Eff. (w)=-†2026.03 | 12.47 | |
| SERQ (GPTQ)Backbone=LLaMA-3.2 1B, #Bits=W4A4, #Eff. (w)=4.242026.03 | 12.52 |