General Language Understanding on tinyBenchmark
77.51Accuracy (ARC)No Steering
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| No SteeringBase Model=Qwen2.5-32B-Instruct2026.02 | 77.51 | 87.75 | 77.12 | 61.49 | 76.3 | 76.03 | — | |
| No SteeringBackbone=Qwen2.5-32B-Instruct, Intervention protocol=None2026.02 | 77.51 | 87.75 | 77.12 | 61.49 | 76.3 | 76.03 | — | |
| CHaRS-PCTBase Model=Qwen2.5-32B-Instruct2026.02 | 75.86 | 86.89 | 77.31 | 61.85 | 77.05 | 75.79 | — | |
| DirAblBase Model=Qwen2.5-32B-Instruct2026.02 | 75.13 | 86.89 | 77.31 | 61.78 | 74.03 | 75.03 | — | |
| CHaRSBase Model=Qwen2.5-32B-Instruct2026.02 | 74.08 | 86.68 | 78.11 | 61.8 | 77.36 | 75.61 | — | |
| No SteeringBase Model=Qwen2.5-14B-Instruct2026.02 | 73.96 | 82.71 | 74.6 | 64.5 | 73.77 | 73.91 | — | |
| DirAblBase Model=Qwen2.5-14B-Instruct2026.02 | 73.96 | 83.14 | 74.87 | 64.73 | 75.93 | 74.53 | — | |
| No SteeringBackbone=Qwen2.5-14B-Instruct, Intervention protocol=None2026.02 | 73.96 | 82.71 | 74.6 | 64.5 | 73.77 | 73.91 | — | |
| OriginalModel Backbone=Qwen2.5-14B Instruct2025.10 | 73.96 | 82.7 | 74.6 | 64.5 | 73.77 | — | 90.12 | |
| OriginalModel Backbone=Gemma2-27B Instruct2025.10 | 73.45 | 83.24 | 76.11 | 61.36 | 75.47 | — | 86.91 | |
| ITIModel Backbone=Qwen2.5-14B Instruct2025.10 | 73.15 | 82.24 | 74.55 | 64.03 | 73.31 | — | 89.27 | |
| PIDModel Backbone=Gemma2-27B Instruct2025.10 | 72.93 | 82.95 | 75.67 | 60.74 | 75.06 | — | 86.6 | |
| ITIModel Backbone=Gemma2-27B Instruct2025.10 | 72.84 | 82.83 | 75.51 | 60.85 | 75.11 | — | 86.38 | |
| DIMModel Backbone=Qwen2.5-14B Instruct2025.10 | 72.74 | 81.94 | 74.3 | 63.01 | 72.93 | — | 87.01 | |
| CHaRS-PCTBase Model=Qwen2.5-14B-Instruct2026.02 | 72.69 | 82.89 | 74.63 | 63.87 | 72.65 | 73.35 | — | |
| CHaRSBase Model=Qwen2.5-14B-Instruct2026.02 | 72.52 | 82.71 | 74.63 | 63.44 | 72.23 | 73.11 | — | |
| RePEModel Backbone=Qwen2.5-14B Instruct2025.10 | 72.4 | 81.52 | 73.9 | 63.2 | 72.6 | — | 86.2 | |
| PID (ours)Model Backbone=Qwen2.5-14B Instruct2025.10 | 72.13 | 82.6 | 74.52 | 63.6 | 73.04 | — | 88.96 | |
| DIMModel Backbone=Gemma2-27B Instruct2025.10 | 72.13 | 81.79 | 74.59 | 59.49 | 74.21 | — | 84.7 | |
| RePEModel Backbone=Gemma2-27B Instruct2025.10 | 71.03 | 80.82 | 73.9 | 58.79 | 73.44 | — | 84 | |
| No SteeringBase Model=Gemma2-9B-Instruct2026.02 | 69.31 | 82.22 | 76.6 | 55.07 | 72.35 | 71.11 | — | |
| CHaRSBase Model=Gemma2-9B-Instruct2026.02 | 69.31 | 81.35 | 74.9 | 51.82 | 71.02 | 69.68 | — | |
| CHaRS-PCTBase Model=Gemma2-9B-Instruct2026.02 | 69.31 | 81.34 | 75.88 | 52.56 | 72.54 | 70.33 | — | |
| No SteeringBackbone=Gemma2-9B-Instruct, Intervention protocol=None2026.02 | 69.31 | 82.22 | 76.6 | 55.07 | 72.35 | 71.11 | — | |
| OriginalModel Backbone=Gemma2-9B Instruct2025.10 | 69.31 | 82.31 | 76.6 | 55.07 | 72.34 | — | 83.19 | |
| DirAblBase Model=Gemma2-9B-Instruct2026.02 | 68.55 | 80.49 | 75.88 | 52.55 | 71.86 | 69.87 | — | |
| No SteeringBase Model=Qwen2.5-7B-Instruct2026.02 | 68.36 | 78.88 | 72.57 | 56.41 | 75.2 | 70.28 | — | |
| No SteeringBackbone=Qwen2.5-7B-Instruct, Intervention protocol=None2026.02 | 68.36 | 78.88 | 72.57 | 56.41 | 75.2 | 70.28 | — | |
| ITIModel Backbone=Gemma2-9B Instruct2025.10 | 68.32 | 81.7 | 75.33 | 53.13 | 71.82 | — | 81.47 | |
| CHaRS-PCTBase Model=Qwen2.5-7B-Instruct2026.02 | 68.28 | 76.77 | 72.45 | 57.32 | 75.28 | 70.02 | — | |
| DIMModel Backbone=Gemma2-9B Instruct2025.10 | 68.21 | 81.45 | 72.29 | 51.86 | 71.51 | — | 80.14 | |
| PID (ours)Model Backbone=Gemma2-9B Instruct2025.10 | 67.91 | 81.59 | 74.89 | 52.49 | 71.42 | — | 79.24 | |
| CHaRSBase Model=Qwen2.5-7B-Instruct2026.02 | 67.71 | 79.58 | 71.22 | 57.35 | 73.59 | 69.89 | — | |
| RePEModel Backbone=Gemma2-9B Instruct2025.10 | 67.5 | 81.2 | 71.1 | 51.1 | 71.15 | — | 79.2 | |
| DirAblBase Model=Qwen2.5-7B-Instruct2026.02 | 67.15 | 76.89 | 72.76 | 57.9 | 76.67 | 70.27 | — | |
| No SteeringBase Model=Llama3.1-8B-Instruct2026.02 | 65.33 | 82.51 | 62.02 | 54.39 | 65.57 | 65.96 | — | |
| No SteeringBackbone=Llama3.1-8B-Instruct, Intervention protocol=None2026.02 | 65.33 | 82.51 | 62.02 | 54.39 | 65.57 | 65.96 | — | |
| OriginalModel Backbone=Llama3.1-8B Instruct2025.10 | 65.33 | 82.51 | 62.02 | 54.39 | 65.56 | — | 63.21 | |
| CHaRS-PCTBase Model=Llama3.1-8B-Instruct2026.02 | 65.12 | 80.46 | 64.05 | 54.53 | 65.52 | 65.94 | — | |
| CHaRSBase Model=Llama3.1-8B-Instruct2026.02 | 64.86 | 80.18 | 63.55 | 54.66 | 67.34 | 66.12 | — | |
| ITIModel Backbone=Llama3.1-8B Instruct2025.10 | 64.26 | 82.01 | 61.37 | 54.33 | 65.21 | — | 61.85 | |
| DirAblBase Model=Llama3.1-8B-Instruct2026.02 | 63.95 | 80.46 | 62.13 | 54.58 | 65.93 | 65.41 | — | |
| CHaRSBase Model=Qwen2.5-3B-Instruct2026.02 | 62.42 | 70.78 | 66.48 | 57.99 | 66.07 | 64.75 | — | |
| PID (ours)Model Backbone=Llama3.1-8B Instruct2025.10 | 62.3 | 81.87 | 61.54 | 54.24 | 64.93 | — | 61.99 | |
| No SteeringBase Model=Qwen2.5-3B-Instruct2026.02 | 62.29 | 73.18 | 68.03 | 56.43 | 70.65 | 66.12 | — | |
| No SteeringBackbone=Qwen2.5-3B-Instruct, Intervention protocol=None2026.02 | 62.29 | 73.18 | 68.03 | 56.43 | 70.65 | 66.12 | — | |
| OriginalModel Backbone=Qwen2.5-3B Instruct2025.10 | 62.29 | 73.18 | 68.03 | 56.43 | 70.65 | — | 17.64 | |
| DIMModel Backbone=Llama3.1-8B Instruct2025.10 | 62.01 | 81.73 | 60.96 | 54.17 | 64.81 | — | 60.57 | |
| DIMModel Backbone=Qwen2.5-3B Instruct2025.10 | 61.95 | 72.4 | 66.11 | 54.95 | 69.85 | — | 14.8 | |
| RePEModel Backbone=Llama3.1-8B Instruct2025.10 | 61.4 | 81.35 | 60.2 | 53.7 | 64.45 | — | 60 | |
| ITIModel Backbone=Qwen2.5-3B Instruct2025.10 | 61.28 | 72.71 | 66.62 | 54.75 | 70.12 | — | 15.57 | |
| PID (ours)Model Backbone=Qwen2.5-3B Instruct2025.10 | 61.2 | 72.59 | 67.29 | 54.1 | 69.72 | — | 16.01 | |
| HiDRABackbone=Qwen2.5-7B-Instruct, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 61.2 | 79.42 | 71.6 | 65.08 | 72.71 | — | — | |
| RePEModel Backbone=Qwen2.5-3B Instruct2025.10 | 61.05 | 72.03 | 65.7 | 54.3 | 69.4 | — | 14.6 | |
| Mean-AcTBackbone=Qwen2.5-7B-Instruct2026.06 | 60.83 | 79.37 | 72.68 | 64.77 | 71.56 | — | — | |
| CHaRSBackbone=Qwen2.5-32B-Instruct, Intervention protocol=ActAdd2026.02 | 60.71 | 75.13 | 76.97 | 54.81 | 73.16 | 68.16 | — | |
| DirAblBase Model=Qwen2.5-3B-Instruct2026.02 | 60.59 | 72.67 | 66.19 | 58.37 | 62.93 | 64.15 | — | |
| CHaRS-PCTBackbone=Qwen2.5-32B-Instruct, Intervention protocol=ActAdd2026.02 | 60.43 | 76.91 | 77.4 | 54.93 | 72.66 | 68.47 | — | |
| CHaRS-PCTBase Model=Qwen2.5-3B-Instruct2026.02 | 59.87 | 74.99 | 66.01 | 58.63 | 59.8 | 63.86 | — | |
| ActAddBackbone=Qwen2.5-32B-Instruct, Intervention protocol=ActAdd2026.02 | 59.72 | 76.86 | 77.4 | 55.04 | 75.06 | 68.82 | — | |
| CHaRSBackbone=Qwen2.5-14B-Instruct, Intervention protocol=ActAdd2026.02 | 59.18 | 74.17 | 63.73 | 56.22 | 70.69 | 64.8 | — | |
| CHaRS-PCTBackbone=Qwen2.5-14B-Instruct, Intervention protocol=ActAdd2026.02 | 58.28 | 73.43 | 64.54 | 55.44 | 71.99 | 64.74 | — | |
| Mean-AcTBackbone=Qwen2.5-3B-Instruct2026.06 | 58.2 | 70.69 | 62.59 | 63.66 | 64.81 | — | — | |
| ActAddBackbone=Qwen2.5-14B-Instruct, Intervention protocol=ActAdd2026.02 | 57.71 | 73.43 | 65.94 | 55.84 | 70.79 | 64.74 | — | |
| CHaRSBackbone=Qwen2.5-7B-Instruct, Intervention protocol=ActAdd2026.02 | 57.7 | 72.44 | 62.45 | 51.95 | 63.75 | 61.66 | — | |
| HiDRABackbone=Qwen2.5-3B-Instruct, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 57.32 | 70.7 | 61.74 | 64.44 | 65.56 | — | — | |
| CHaRS-PCTBackbone=Qwen2.5-7B-Instruct, Intervention protocol=ActAdd2026.02 | 57.02 | 72.44 | 64.19 | 51.96 | 65.47 | 62.22 | — | |
| ActAddBackbone=Qwen2.5-7B-Instruct, Intervention protocol=ActAdd2026.02 | 56.96 | 72.77 | 64.75 | 51.32 | 65.71 | 62.3 | — | |
| CHaRSBase Model=Llama3.2-3B-Instruct2026.02 | 56.44 | 76.9 | 60.68 | 50.91 | 57.56 | 60.5 | — | |
| No SteeringBase Model=Llama3.2-3B-Instruct2026.02 | 55.86 | 75.92 | 63.48 | 50.19 | 58.64 | 60.82 | — | |
| No SteeringBackbone=Llama3.2-3B-Instruct, Intervention protocol=None2026.02 | 55.86 | 75.92 | 63.48 | 50.19 | 58.64 | 60.82 | — | |
| DirAblBase Model=Llama3.2-3B-Instruct2026.02 | 55.51 | 77.19 | 61.21 | 51.09 | 59.33 | 60.87 | — | |
| CHaRS-PCTBase Model=Llama3.2-3B-Instruct2026.02 | 55.15 | 78.94 | 63.15 | 50.58 | 58.33 | 61.23 | — | |
| CHaRS-PCTBackbone=Llama3.1-8B-Instruct, Intervention protocol=ActAdd2026.02 | 53.52 | 79.44 | 57.23 | 47.3 | 69.55 | 61.41 | — | |
| ActAddBackbone=Llama3.1-8B-Instruct, Intervention protocol=ActAdd2026.02 | 52.68 | 79.44 | 56.78 | 47.44 | 68.33 | 60.93 | — | |
| CHaRSBackbone=Llama3.1-8B-Instruct, Intervention protocol=ActAdd2026.02 | 51.71 | 79.9 | 55.61 | 47.45 | 68.79 | 60.69 | — | |
| ActAddBackbone=Gemma2-9B-Instruct, Intervention protocol=ActAdd2026.02 | 50.31 | 62.94 | 41.9 | 42.7 | 59.08 | 51.39 | — | |
| CHaRSBackbone=Gemma2-9B-Instruct, Intervention protocol=ActAdd2026.02 | 50.31 | 62.88 | 46.5 | 42.79 | 59.21 | 52.34 | — | |
| CHaRS-PCTBackbone=Llama3.2-3B-Instruct, Intervention protocol=ActAdd2026.02 | 50.21 | 61.7 | 58.53 | 40.47 | 54.14 | 53.01 | — | |
| CHaRS-PCTBackbone=Gemma2-9B-Instruct, Intervention protocol=ActAdd2026.02 | 50.03 | 61.7 | 43.34 | 42.81 | 60.21 | 51.62 | — | |
| ActAddBackbone=Llama3.2-3B-Instruct, Intervention protocol=ActAdd2026.02 | 49.68 | 63.54 | 57.33 | 40.72 | 56.75 | 53.6 | — | |
| CHaRSBackbone=Llama3.2-3B-Instruct, Intervention protocol=ActAdd2026.02 | 48.64 | 63.54 | 57.02 | 40.47 | 55.7 | 53.07 | — | |
| HiDRABackbone=Gemma2-9B-IT, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 46.84 | 69.21 | 60.92 | 51.97 | 69.38 | — | — | |
| Mean-AcTBackbone=Gemma2-9B-IT2026.06 | 46.6 | 70.66 | 62.1 | 50.84 | 64.59 | — | — | |
| CHaRS-PCTBackbone=Qwen2.5-3B-Instruct, Intervention protocol=ActAdd2026.02 | 46.18 | 62.67 | 61.97 | 45.55 | 64.12 | 56.1 | — | |
| ActAddBackbone=Qwen2.5-3B-Instruct, Intervention protocol=ActAdd2026.02 | 45.99 | 62.67 | 61.36 | 45.96 | 64.3 | 56.06 | — | |
| CHaRSBackbone=Qwen2.5-3B-Instruct, Intervention protocol=ActAdd2026.02 | 45.1 | 63.81 | 61.82 | 45.73 | 64.5 | 56.19 | — | |
| HiDRABackbone=Llama3.2-3B-Instruct, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 44.02 | 66.79 | 49.75 | 46.8 | 58.19 | — | — | |
| Mean-AcTBackbone=Llama3.2-3B-Instruct2026.06 | 43.82 | 67.9 | 49.43 | 47.25 | 56.18 | — | — | |
| HiDRABackbone=Llama3.2-1B-Instruct, projected dimensions (m)=8192, LeakyReLU slope={0.5, 0.7}2026.06 | 36.71 | 63.25 | 44.07 | 48.85 | 56.51 | — | — | |
| Mean-AcTBackbone=Llama3.2-1B-Instruct2026.06 | 36.41 | 63.14 | 41.61 | 48.49 | 58.85 | — | — |