Linguistic Acceptability on CoLA (Accuracy)
85.91Accuracy (CoLA)LoRA-Mixer (ours)
Evaluation results
| Method | Links | ||
|---|---|---|---|
| LoRA-Mixer (ours)Base Model=Falcon-Mamba (7B)2025.06 | 85.91 | — | |
| LoRABase Model=Falcon-Mamba (7B)2025.06 | 85.62 | — | |
| MOLEBase Model=Falcon-Mamba (7B)2025.06 | 84.77 | — | |
| UnpatchedAttack=No attack2026.06 | 83 | 1 | |
| UnpatchedAttack=Sleeper2026.06 | 83 | 99 | |
| LoRA-Mixer(ours)Base Model=LLaMA-3 (8B)2025.06 | 82.22 | — | |
| LoRA-Mixer (ours)Base Model=Mistral (7B)2025.06 | 82.17 | — | |
| UnpatchedAttack=UJB2026.06 | 82 | 97 | |
| UnpatchedAttack=BadNets2026.06 | 82 | 96 | |
| UnpatchedAttack=CBA2026.06 | 82 | 97 | |
| UnpatchedAttack=Sleeper2026.06 | 82 | 98 | |
| PatcherAttack=No attack2026.06 | 82 | 1 | |
| UnpatchedAttack=UJB2026.06 | 82 | 98 | |
| LoRABase Model=LLaMA-3 (8B)2025.06 | 81.5 | — | |
| MOLEBase Model=LLaMA-3 (8B)2025.06 | 81.37 | — | |
| LoRA-MixerBackbone=LLaMA3-8B, Adapter=Q/K/V/O, Routing Strategy=RSL2025.06 | 81.17 | — | |
| LoRAHubBase Model=Falcon-Mamba (7B)2025.06 | 81.11 | — | |
| UnpatchedAttack=No attack2026.06 | 81 | 4 | |
| FTAttack=UJB2026.06 | 81 | 96 | |
| FTAttack=BadNets2026.06 | 81 | 96 | |
| FTAttack=CBA2026.06 | 81 | 97 | |
| FTℓ2Attack=BadNets2026.06 | 81 | 96 | |
| FTℓ∞Attack=BadNets2026.06 | 81 | 95 | |
| FTℓ∞Attack=CBA2026.06 | 81 | 96 | |
| FinePruningAttack=BadNets2026.06 | 81 | 96 | |
| BAERASERAttack=UJB2026.06 | 81 | 97 | |
| MudjackingAttack=No attack2026.06 | 81 | 2 | |
| MudjackingAttack=CBA2026.06 | 81 | 85 | |
| OneShotAttack=UJB2026.06 | 81 | 96 | |
| OneShotAttack=CBA2026.06 | 81 | 95 | |
| MENDAttack=BadNets2026.06 | 81 | 96 | |
| MENDAttack=CBA2026.06 | 81 | 97 | |
| MENDAttack=Sleeper2026.06 | 81 | 97 | |
| ROMEAttack=CBA2026.06 | 81 | 95 | |
| MixLoRABase Model=LLaMA-3 (8B)2025.06 | 80.67 | — | |
| UnpatchedAttack=POLISHED2026.06 | 80 | 97 | |
| FTAttack=No attack2026.06 | 80 | 6 | |
| FTℓ2Attack=Sleeper2026.06 | 80 | 98 | |
| FTℓ∞Attack=No attack2026.06 | 80 | 4 | |
| FTℓ∞Attack=UJB2026.06 | 80 | 95 | |
| FinePruningAttack=No attack2026.06 | 80 | 6 | |
| BAERASERAttack=No attack2026.06 | 80 | 2 | |
| MudjackingAttack=Sleeper2026.06 | 80 | 68 | |
| SPPAttack=UJB2026.06 | 80 | 97 | |
| OneShotAttack=BadNets2026.06 | 80 | 96 | |
| MENDAttack=No attack2026.06 | 80 | 5 | |
| MENDAttack=UJB2026.06 | 80 | 95 | |
| ROMEAttack=BadNets2026.06 | 80 | 94 | |
| PatcherAttack=No attack2026.06 | 80 | 2 | |
| PatcherAttack=UJB2026.06 | 80 | 1 | |
| PatcherAttack=BadNets2026.06 | 80 | 1 | |
| PatcherAttack=CBA2026.06 | 80 | 1 | |
| PatcherAttack=Sleeper2026.06 | 80 | 2 | |
| PatcherAttack=UJB2026.06 | 80 | 1 | |
| PatcherAttack=Sleeper2026.06 | 80 | 0 | |
| LoRAHubBase Model=LLaMA-3 (8B)2025.06 | 79.84 | — | |
| LoRABase Model=Mistral (7B)2025.06 | 79.19 | — | |
| BaseBackbone=LLaMA3-8B, Adapter=Q/K/V/O2025.06 | 79.14 | — | |
| FTAttack=POLISHED2026.06 | 79 | 96 | |
| FTℓ2Attack=No attack2026.06 | 79 | 4 | |
| FTℓ2Attack=POLISHED2026.06 | 79 | 97 | |
| FTℓ∞Attack=POLISHED2026.06 | 79 | 96 | |
| BAERASERAttack=POLISHED2026.06 | 79 | 97 | |
| BAERASERAttack=Sleeper2026.06 | 79 | 95 | |
| MudjackingAttack=UJB2026.06 | 79 | 97 | |
| MudjackingAttack=BadNets2026.06 | 79 | 96 | |
| SPPAttack=No attack2026.06 | 79 | 7 | |
| SPPAttack=CBA2026.06 | 79 | 96 | |
| OneShotAttack=No attack2026.06 | 79 | 3 | |
| MENDAttack=POLISHED2026.06 | 79 | 97 | |
| ROMEAttack=No attack2026.06 | 79 | 3 | |
| ROMEAttack=Sleeper2026.06 | 79 | 98 | |
| MudjackingAttack=Sleeper2026.06 | 79 | 44 | |
| MixLoRABase Model=Mistral (7B)2025.06 | 78.61 | — | |
| MOLEBase Model=Mistral (7B)2025.06 | 78.51 | — | |
| PhatgooseBackbone=LLaMA3-8B, Adapter=Q/K/V/O, Routing Strategy=sigmoid gating based on cosine similarity2025.06 | 78.27 | — | |
| Hard-Routed MoR-LoRABackbone=Llama-8B, # Trainable Parameters=≈ 109M2026.06 | 78.09 | — | |
| FTℓ2Attack=UJB2026.06 | 78 | 97 | |
| FinePruningAttack=CBA2026.06 | 78 | 96 | |
| FinePruningAttack=POLISHED2026.06 | 78 | 97 | |
| FinePruningAttack=Sleeper2026.06 | 78 | 98 | |
| BAERASERAttack=CBA2026.06 | 78 | 96 | |
| MudjackingAttack=POLISHED2026.06 | 78 | 97 | |
| SPPAttack=Sleeper2026.06 | 78 | 98 | |
| ROMEAttack=POLISHED2026.06 | 78 | 95 | |
| PatcherAttack=POLISHED2026.06 | 78 | 3 | |
| LoRAMixer TopK=2 NormalizedBackbone=Llama-8B, # Trainable Parameters=≈ 1.133B2026.06 | 77.08 | — | |
| FTℓ2Attack=CBA2026.06 | 77 | 97 | |
| FTℓ∞Attack=Sleeper2026.06 | 77 | 97 | |
| FinePruningAttack=UJB2026.06 | 77 | 97 | |
| BAERASERAttack=BadNets2026.06 | 77 | 96 | |
| SPPAttack=BadNets2026.06 | 77 | 96 | |
| SPPAttack=POLISHED2026.06 | 77 | 97 | |
| OneShotAttack=Sleeper2026.06 | 77 | 96 | |
| ROMEAttack=UJB2026.06 | 77 | 93 | |
| LoRAMixer TopK=1Backbone=Llama-8B, # Trainable Parameters=≈ 1.133B2026.06 | 76.99 | — | |
| Hard-Routed MoR-LoRABackbone=Llama-3B, # Trainable Parameters=≈ 73M2026.06 | 76.7 | — | |
| OneShotAttack=POLISHED2026.06 | 76 | 97 | |
| LoRAHubBase Model=Mistral (7B)2025.06 | 75.73 | — | |
| LoRAMixer TopK=2 NormalizedBackbone=Llama-3B, # Trainable Parameters=≈ 606M2026.06 | 75.46 | — |