Hallucination Reduction on CAA Hallucination Benchmark Multiple-Choice
86.65Alignment ProbabilityINNSTEER
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| INNSTEERModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 86.65 | 2.98 | |
| INNSTEERModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 83.32 | 4.58 | |
| ODEModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 77.6 | 3.03 | |
| TFModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 75.3 | 3.04 | |
| ITIModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 74.2 | 3.03 | |
| LAcTModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 73.8 | 3.03 | |
| ODEModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 73.5 | 4.53 | |
| HPRModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 73.1 | 3.02 | |
| INNSTEERModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 72.41 | 3.2 | |
| TFModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 71.2 | 4.55 | |
| ITIModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 70.4 | 4.52 | |
| LAcTModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 70 | 4.52 | |
| HPRModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 69.1 | 4.5 | |
| CAAModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 69.02 | 3 | |
| RECModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 68.5 | 3 | |
| INNSTEERModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 68.28 | 3.18 | |
| BaseModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 68.05 | 2.94 | |
| BaseModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 68.05 | 2.94 | |
| PoEModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 67.6 | 2.99 | |
| PoDModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 67.4 | 2.98 | |
| PoEModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 67.4 | 2.99 | |
| PoDModel Backbone=LLaMA3-8B, Steering Direction=Positive2026.06 | 67.2 | 2.98 | |
| RECModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 67.1 | 3 | |
| CAAModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 66.9 | 3 | |
| CAAModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 65.58 | 4.48 | |
| HPRModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 65.4 | 3.02 | |
| RECModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 65.1 | 4.49 | |
| BaseModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 64.61 | 3.95 | |
| BaseModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 64.61 | 4.45 | |
| ITIModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 64.5 | 3.03 | |
| LAcTModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 64.2 | 3.03 | |
| ODEModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 64.1 | 3.12 | |
| PoEModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 64.05 | 4.47 | |
| TFModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 63.8 | 3.04 | |
| PoDModel Backbone=LLaMA3-3B, Steering Direction=Positive2026.06 | 63.8 | 4.47 | |
| PoEModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 63.4 | 3.97 | |
| PoDModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 63.2 | 3.97 | |
| ODEModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 62.5 | 3.03 | |
| RECModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 62.4 | 3.99 | |
| TFModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 62.4 | 3.13 | |
| ITIModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 61.8 | 3.12 | |
| CAAModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 61.78 | 3.98 | |
| LAcTModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 61.5 | 3.11 | |
| INNSTEERModel=LLaMA3-8B, Steering direction=Negative steering2026.06 | 61.1 | 2.98 | |
| INNSTEERModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 60.79 | 4.07 | |
| HPRModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 60.7 | 4.01 | |
| HPRModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 60.7 | 3.1 | |
| ITIModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 59.8 | 4.02 | |
| LAcTModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 59.5 | 4.02 | |
| TFModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 58.9 | 4.05 | |
| CAAModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 58.12 | 3.08 | |
| RECModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 57.9 | 3.07 | |
| BaseModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 57.36 | 3.05 | |
| ODEModel=LLaMA3-3B, Steering direction=Negative steering2026.06 | 57.2 | 4.03 | |
| PoEModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 57.05 | 3.05 | |
| CAAModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 56.9 | 3.08 | |
| PoDModel Backbone=Qwen2.5-3B, Steering Direction=Positive2026.06 | 56.9 | 3.04 | |
| RECModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 56.8 | 3.07 | |
| BaseModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 56.79 | 3.05 | |
| PoEModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 56.6 | 3.05 | |
| PoDModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 56.5 | 3.04 | |
| ODEModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 55.9 | 3.18 | |
| HPRModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 55.6 | 3.1 | |
| ITIModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 54.9 | 3.12 | |
| LAcTModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 54.7 | 3.11 | |
| TFModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 54.4 | 3.13 | |
| TFModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 53.8 | 3.19 | |
| ODEModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 53.6 | 3.12 | |
| ITIModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 52.9 | 3.18 | |
| LAcTModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 52.4 | 3.17 | |
| HPRModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 51.8 | 3.16 | |
| INNSTEERModel=Qwen2.5-3B, Steering direction=Negative steering2026.06 | 50.27 | 3.18 | |
| CAAModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 47.59 | 3.12 | |
| RECModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 47.1 | 3.11 | |
| INNSTEERModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 46.99 | 3.25 | |
| BaseModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 46.3 | 3.1 | |
| BaseModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 46.3 | 3.1 | |
| PoEModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 46 | 3.1 | |
| PoDModel Backbone=Qwen2.5-7B, Steering Direction=Positive2026.06 | 45.8 | 3.09 | |
| PoEModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 45.3 | 3.1 | |
| PoDModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 45.1 | 3.09 | |
| ODEModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 44.8 | 3.22 | |
| RECModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 44.2 | 3.11 | |
| CAAModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 43.92 | 3.12 | |
| TFModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 43.4 | 3.23 | |
| HPRModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 42.8 | 3.18 | |
| ITIModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 42.8 | 3.22 | |
| LAcTModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 42.6 | 3.21 | |
| HPRModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 42.2 | 3.2 | |
| ITIModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 41.7 | 3.18 | |
| LAcTModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 41.3 | 3.17 | |
| CAAModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 41.16 | 3.18 | |
| TFModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 40.9 | 3.19 | |
| RECModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 40.8 | 3.17 | |
| BaseModel=Qwen2.5-32B, Steering direction=Negative steering2026.06 | 40.42 | 3.15 | |
| BaseModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 40.42 | 3.15 | |
| PoEModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 40.1 | 3.15 | |
| PoDModel Backbone=Qwen2.5-32B, Steering Direction=Positive2026.06 | 39.9 | 3.14 | |
| ODEModel=Qwen2.5-7B, Steering direction=Negative steering2026.06 | 39.8 | 3.18 | |
| PoEModel=Qwen2.5-32B, Steering direction=Negative steering2026.06 | 39.1 | 3.15 |