Toxicity Detection on ToxiGen
84.23ScoreTaT (Disp.)
Evaluation Results
| Method | Links | |
|---|---|---|
| TaT (Disp.)Backbone=Llama3.1-8b2026.03 | 84.23 | |
| Trajectory (Raw)Backbone=Qwen2.5-14B2026.03 | 83.48 | |
| TaT (Disp.)Backbone=Qwen3-30B MoE2026.03 | 82.34 | |
| TaT (Disp.)Backbone=Qwen2.5-14B2026.03 | 82.28 | |
| Trajectory (Raw)Backbone=Llama3.1-8b2026.03 | 81.99 | |
| Trajectory (Raw)Backbone=Qwen3-30B MoE2026.03 | 81.93 | |
| OLMo (base)Parameter Count=7B2024.02 | 81.4 | |
| TaT (Disp.)Backbone=Qwen3-32B2026.03 | 81.4 | |
| Trajectory (Raw)Backbone=Qwen3-32B2026.03 | 80.22 | |
| Linear ProbeBackbone=Llama3.1-8b2026.03 | 79.62 | |
| DOPA*Backbone=Qwen3-8B2026.04 | 79 | |
| DICLBackbone=Qwen3-8B2026.04 | 77.53 | |
| DrICLBackbone=Qwen3-8B2026.04 | 77.43 | |
| Linear ProbeBackbone=Qwen3-30B MoE2026.03 | 75.16 | |
| Linear ProbeBackbone=Qwen2.5-14B2026.03 | 72.58 | |
| Falcon InstructParameter Count=7B2024.02 | 70.7 | |
| RewriteBackbone=Qwen3-8B2026.04 | 70.5 | |
| KNNBackbone=Qwen3-8B2026.04 | 68.07 | |
| DrICLBackbone=LLaMA3.1-8B2026.04 | 67.73 | |
| DOPA*Backbone=LLaMA3.1-8B2026.04 | 67.43 | |
| RandomBackbone=Qwen3-1.7B2026.04 | 67.2 | |
| RandomBackbone=LLaMA3.1-8B2026.04 | 66.4 | |
| DrICLBackbone=Qwen3-1.7B2026.04 | 66.37 | |
| DICLBackbone=LLaMA3.2-3B2026.04 | 66 | |
| DICLBackbone=LLaMA3.1-8B2026.04 | 65.97 | |
| InfICLBackbone=Qwen3-8B2026.04 | 65.97 | |
| DOPA*Backbone=Gemma2-2B2026.04 | 65.73 | |
| KNNBackbone=Qwen3-1.7B2026.04 | 65.73 | |
| DOPA*Backbone=Qwen3-1.7B2026.04 | 65.73 | |
| RandomBackbone=LLaMA3.2-3B2026.04 | 65.7 | |
| RandomBackbone=Qwen3-8B2026.04 | 65.63 | |
| InfICLBackbone=LLaMA3.2-3B2026.04 | 65.57 | |
| KNNBackbone=LLaMA3.1-8B2026.04 | 65.5 | |
| DOPA*Backbone=LLaMA3.2-3B2026.04 | 65.3 | |
| DICLBackbone=Gemma2-2B2026.04 | 65.3 | |
| InfICLBackbone=Qwen3-1.7B2026.04 | 65.23 | |
| DrICLBackbone=LLaMA3.2-3B2026.04 | 64.8 | |
| DICLBackbone=Qwen3-1.7B2026.04 | 64.45 | |
| RandomBackbone=Gemma2-2B2026.04 | 63.93 | |
| KNNBackbone=Gemma2-2B2026.04 | 63.87 | |
| RewriteBackbone=LLaMA3.1-8B2026.04 | 63.87 | |
| InfICLBackbone=LLaMA3.1-8B2026.04 | 63.77 | |
| RewriteBackbone=LLaMA3.2-3B2026.04 | 62.7 | |
| Linear ProbeBackbone=Qwen3-32B2026.03 | 62.24 | |
| RewriteBackbone=Qwen3-1.7B2026.04 | 61.97 | |
| RewriteBackbone=Gemma2-2B2026.04 | 61.84 | |
| DrICLBackbone=Gemma2-2B2026.04 | 61.17 | |
| InfICLBackbone=Gemma2-2B2026.04 | 59.9 | |
| MAT-STEERBackbone=Llama-3.1-8B-Chat2025.02 | 57.82 | |
| LITOBackbone=Llama-3.1-8B-Chat2025.02 | 56.94 | |
| SFTBackbone=Llama-3.1-8B-Chat2025.02 | 56.45 | |
| DPOBackbone=Llama-3.1-8B-Chat2025.02 | 56.11 | |
| GRPOBackbone=Qwen3-8B, Scenario=Strategic Deception2026.03 | 55.96 | |
| SARBackbone=Qwen3-8B, Scenario=Strategic Deception2026.03 | 55.74 | |
| CoT MonitorBackbone=Llama-3.1-8B, Scenario=Strategic Deception2026.03 | 55.53 | |
| ICLBackbone=Llama-3.1-8B-Chat2025.02 | 55.37 | |
| CoT MonitorBackbone=Qwen3-8B, Scenario=Strategic Deception2026.03 | 54.89 | |
| RAdaptBackbone=Llama-3.1-8B-Chat2025.02 | 54.74 | |
| MergeBackbone=Llama-3.1-8B-Chat2025.02 | 54.06 | |
| GRPOBackbone=Llama-3.1-8B, Scenario=Strategic Deception2026.03 | 53.51 | |
| Llama-3.1-8BScenario=Strategic Deception2026.03 | 53.3 | |
| Llama-3.1-8BScenario=Sycophancy2026.03 | 53.3 | |
| ICVBackbone=Llama-3.1-8B-Chat2025.02 | 53.03 | |
| SARBackbone=Llama-3.1-8B, Scenario=Sycophancy2026.03 | 52.87 | |
| GRPOBackbone=Llama-3.1-8B, Scenario=Sycophancy2026.03 | 52.66 | |
| CoT MonitorBackbone=Llama-3.1-8B, Scenario=Sycophancy2026.03 | 52.55 | |
| NL-ITIBackbone=Llama-3.1-8B-Chat2025.02 | 52.37 | |
| SARBackbone=Qwen3-8B, Scenario=Sycophancy2026.03 | 52.13 | |
| ITIBackbone=Llama-3.1-8B-Chat2025.02 | 52.06 | |
| GRPOBackbone=Qwen3-8B, Scenario=Sycophancy2026.03 | 51.6 | |
| CoT MonitorBackbone=Qwen3-8B, Scenario=Sycophancy2026.03 | 51.6 | |
| Qwen3-8BScenario=Strategic Deception2026.03 | 51.28 | |
| Qwen3-8BScenario=Sycophancy2026.03 | 51.28 | |
| KNNBackbone=LLaMA3.2-3B2026.04 | 51.03 | |
| RandomBackbone=GPT2-xl2026.04 | 50.6 | |
| DOPA*Backbone=GPT2-xl2026.04 | 50.5 | |
| InfICLBackbone=GPT2-xl2026.04 | 50.17 | |
| Llama-3.1-ChatBackbone=Llama-3.1-8B-Chat2025.02 | 49.97 | |
| BaselineBackbone=LLaMA2-7B-CHAT2025.12 | 49.71 | |
| CAABackbone=LLaMA2-7B-CHAT2025.12 | 49.71 | |
| RewriteBackbone=GPT2-xl2026.04 | 49.37 | |
| KNNBackbone=GPT2-xl2026.04 | 48.5 | |
| DrICLBackbone=GPT2-xl2026.04 | 48.23 | |
| SARBackbone=Llama-3.1-8B, Scenario=Strategic Deception2026.03 | 48.09 | |
| LLM-CASBackbone=LLaMA2-7B-CHAT2025.12 | 47.63 | |
| RPJ-INCITE ChatParameter Count=7B2024.02 | 46.4 | |
| DICLBackbone=GPT2-xl2026.04 | 46.37 | |
| ITIBackbone=LLaMA2-7B-CHAT2025.12 | 45.27 | |
| SADIBackbone=LLaMA2-7B-CHAT2025.12 | 17.14 | |
| OLMo+SFTParameter Count=7B2024.02 | 14.4 | |
| TÜLU 2Parameter Count=7B2024.02 | 7 | |
| OLMo+SFT+DPOParameter Count=7B2024.02 | 1.7 | |
| TÜLU 2+DPOParameter Count=7B2024.02 | 0.5 | |
| MPT ChatParameter Count=7B2024.02 | 0.1 | |
| Llama-2-ChatParameter Count=7B2024.02 | 0 |