| Task Name | Dataset Name | SOTA Result | Trend | |
|---|---|---|---|---|
| Circuit Discovery Evaluation | Gemma-2-2B | Clarity82 | 70 | |
| Automated Interpretability Evaluation | Gemma-2-2B | Clarity80 | 50 | |
| Watermarking Attack Robustness | Gemma 9B v2 (test) | TPR100 | 49 | |
| Negative Sentiment Backdoor Detection | Gemma 2 9B | Attack Success Rate (ASR)0 | 48 | |
| Refusal Backdoor Detection | Gemma-2-9B | ASR0 | 42 | |
| Decoder Reconstruction | GEMMA-3-4B-IT | ROUGE-1 Score99.1 | 40 | |
| Decoder Reconstruction | Gemma 4B IT 3 | ROUGE-198.6 | 24 | |
| Model Steering | Gemma 2 2B Steering Evaluation Set | Granularity1.2961 | 20 | |
| Sparse Autoencoder Evaluation | Gemma-2-2B activations | L0 Count320 | 20 | |
| Jailbreak Attack | Gemma 4B 3 | NR66 | 20 | |
| Jailbreak attack | Gemma-7b five finetuned variants | Average ASR66.2 | 16 | |
| Content Injection | Gemma-2B | Keyword Occurrence (%)92.9 | 15 | |
| Jailbreak Attack | gemma-7b v1 (pretrained) | ASR6 | 13 | |
| Over-refusal Mitigation | Gemma-2B Over-refusal | Informative Refusal Rate1.6 | 12 | |
| LLM Alignment | Gemma-3-4B | Win Rate94.33 | 12 | |
| Text Quality Assessment | Gemma-7B Unmodified watermarked text quality | Perplexity (PPL)12.188 | 11 | |
| Text Watermark Detection | Gemma-7B Post-translation to French human-authored texts | TPR @ 1% FPR59.2 | 11 | |
| Text Watermark Detection | Gemma-7B Post-translation to German human-authored texts | TPR @ 1% FPR53.4 | 11 | |
| Text Watermark Detection | Gemma-7B Post-paraphrasing human-authored texts | TPR @ 1% FPR84 | 11 | |
| Text Watermark Detection | Gemma-7B Unmodified human-authored texts | TPR @ 1% FPR100 | 11 | |
| LLM fingerprinting | Gemma 2 2B | AUC1 | 10 | |
| Language Modeling | Gemma 3 | Accuracy47.06 | 10 | |
| SAE Feature Attribution | Gemma-3-4B 16K SAE | Input Success Rate14.7 | 9 | |
| SAE Feature Attribution | Gemma-3-1B 16K SAE | Input Attribution14.71 | 9 | |
| SAE Feature Attribution | Gemma-3-270M 16K SAE | Input Success Rate (%)17.3 | 9 |