Steer2Edit: From Activation Steering to Component-Level Editing
Organizations: University of California, San Diego
Abstract
Steering methods influence Large Language Model behavior by identifying semantic directions in hidden representations, but are typically realized through inference-time activation interventions that apply a fixed, global modification to the model's internal states. While effective, such interventions often induce unfavorable attribute-utility trade-offs under strong control, as they ignore the fact that many behaviors are governed by a small and heterogeneous subset of model components. We propose Steer2Edit, a theoretically grounded, training-free framework that transforms steering vectors from inference-time control signals into diagnostic signals for component-level rank-1 weight editing. Instead of uniformly injecting a steering direction during generation, Steer2Edit selectively redistributes behavioral influence across individual attention heads and MLP neurons, yielding interpretable edits that preserve the standard forward pass and remain compatible with optimized parallel inference. Across safety alignment, hallucination mitigation, and reasoning efficiency, Steer2Edit consistently achieves more favorable attribute-utility trade-offs: at matched downstream performance, it improves safety by up to 17.2%, increases truthfulness by 9.8%, and reduces reasoning length by 12.2% on average. Overall, Steer2Edit provides a principled bridge between representation steering and weight editing by translating steering signals into interpretable, training-free parameter updates. Our code is available at https://github.com/Trustworthy-ML-Lab/Steer2Edit
Figures & tables
| Method | Training- free | Superior attribute–utility trade-off | Zero inference overhead | Standard forward pass | Component interpretability |
| RepEngineering ( Zou et al., 2023a ) | ✓ | ✗ | ✗ | ✗ | ✗ |
| CAA ( Panickssery et al., 2023 ) | ✓ | ✗ | ✗ | ✗ | ✗ |
| ODESteer ( Zhao et al., 2026 ) | ✗ | ✗ | ✗ | ✗ | ✗ |
| AlphaEdit ( Fang et al., 2024 ) | ✓ | ✗ | ✓ | ✓ | ✗ |
| SFT / LoRA | ✗ | ✗ | ✓ | ✓ | ✗ |
| Steer2Edit (Ours) | ✓ | ✓ | ✓ | ✓ | ✓ |
Appendix figures & tables28 assets
Supplementary material from the paper’s appendix.
Appendix
| Setting / Model | |||
| Safety Alignment | |||
| LLaMA-2-7B-Chat | (step = 0.02) | (step = 0.05) | (step = 0.05) |
| Mistral-7B-Instruct-v0.2 | (step = 0.02) | (step = 0.05) | (step = 0.05) |
| Truthfulness | |||
| Gemma-2-2B-IT | (step = 0.05) | negligible | (step = 0.05) |
| LLaMA-3-8B-Instruct | (step = 0.01) | (step = 0.05) | (step = 0.10) |
| Ablation Category | Variant | Definition and Hypothesis Tested |
| Input | Sets , where . Tests if data statistics alone suffice without directional sensitivity. | |
| Sets , where is the top right singular vector of . Tests if intrinsic weight directions suffice. | ||
| Score | Sets (unnormalized dot product), removing the normalization by . Tests the effect of component-output normalization. | |
| Magnitude | Selects top- components (hard threshold) matching Elastic-Net sparsity. Tests if sparsity alone explains gains. | |
| Uses Elastic-Net with (Ridge regularization), resulting in dense edits. Tests if dense edits can preserve utility. |
| Category | Variant | Normalized Attribute Score | Normalized Utility Score | Overall Average | Attr Util | |||||
| Safety | Truth | Effic. | Safety | Truth | Effic. | Attribute | Utility | |||
| Full Method | Steer2Edit | 0.807 | 0.550 | 1.000 | 0.341 | 0.536 | 0.755 | 0.786 | 0.544 | 0.427 |
| Input | 0.827 | 0.531 | 0.926 | 0.279 | 0.350 | 0.589 | 0.761 | 0.406 | 0.309 | |
| 0.536 | 0.474 | 0.757 | 0.400 | 0.572 | 0.805 | 0.589 | 0.592 | 0.349 | ||
| Score | 0.102 | 0.509 | 0.628 | 0.047 | 0.232 | 0.003 | 0.413 | 0.094 | 0.039 | |
| Magnitude | 0.506 | 0.546 | 0.767 | 0.321 | 0.423 | 0.596 | 0.606 | 0.447 | 0.271 | |
| Model | Variant | S U | Avg | Avg | Safety (Refusal Rate) | Utility (Accuracy) | |||
| Safety | Util | GCG | ADV-LLM | CommonsenseQA | Code-MMLU | GSM8K | |||
| Llama2- 7B-Chat | Steer2Edit | 24.68 | 88.15 | 28.00 | 95.10 | 81.20 | 45.01 | 20.06 | 18.93 |
| 19.12 | 87.90 | 21.75 | 83.60 | 92.20 | 30.14 | 19.57 | 15.53 | ||
| 22.82 | 66.35 | 34.39 | 90.80 | 41.90 | 50.77 | 30.49 | 21.90 | ||
| 0.69 | 15.35 | 4.49 | 23.30 | 7.40 | 6.45 | 5.37 | 1.65 | ||
| 7.66 | 41.25 | 18.58 | 36.40 | 46.10 | 27.26 | 25.43 | 3.05 | ||
| Model | Method | T U | Avg | Avg | Attribute | Utility (Accuracy) | ||
| Truth | Util | TruthfulQA | CommonsenseQA | Code-MMLU | GSM8K | |||
| Gemma2 2B-IT | Steer2Edit | 25.30 | 56.97 | 44.41 | 56.97 | 69.00 | 37.07 | 27.16 |
| 24.93 | 55.75 | 44.71 | 55.75 | 68.59 | 38.29 | 27.24 | ||
| 27.26 | 55.26 | 49.33 | 55.26 | 68.12 | 39.39 | 40.48 | ||
| 1.16 | 58.68 | 1.98 | 58.68 | 2.83 | 2.20 | 0.90 | ||
| 15.33 | 57.95 | 26.46 | 57.95 | 58.29 | 16.04 | 5.06 | ||
| Model | Variant | U / L | Avg | Avg | Reasoning Length (Lower is Better) | Utility (Higher is Better) | ||||||
| Len | Util | MATH-500 | GPQA | Code-MMLU | GSM8K | MATH-500 | GPQA | Code-MMLU | GSM8K | |||
| Qwen3 4B-Thinking | Steer2Edit | 2.28 | 3467 | 78.95 | 4136 | 5299 | 3433 | 1000 | 92.2 | 44.4 | 87.6 | 91.5 |
| 1.56 | 3266 | 51.05 | 3188 | 2728 | 6129 | 1018 | 53.2 | 10.3 | 78.1 | 62.7 | ||
| 1.64 | 5351 | 87.99 | 6389 | 8780 | 4622 | 1613 | 96.9 | 67.5 | 93.9 | 93.6 | ||
| 0.62 | 21 | 0.13 | 23 | 33 | 10 | 18 | 0.3 | 0.0 | 0.0 | 0.2 | ||
| 0.73 | 6506 | 47.61 | 8646 | 5202 | 4698 | 7477 | 44.7 | 36.0 | 65.7 | 44.1 | ||
| Method | GCG Refusal | AdvLLM Refusal | Avg Safety | Avg Utility |
| Original | 70.0% | 29.8% | 49.9% | 35.5% |
| AlphaEdit | 69.6% | 30.9% | 50.3% | 35.0% |
| Steer2Edit | 95.1% | 81.2% | 88.2% | 28.0% |
| Method | TruthfulQA | Expert-FACTOR | Avg Utility |
| Original | 40.3% | 41.5% | 65.8% |
| AlphaEdit | 39.9% | 40.7% | 66.3% |
| Steer2Edit | 53.1% | 54.2% | 62.7% |
| GSM8K | MATH-500 | GPQA | CodeMMLU | Average | ||||||
| Method | Acc | Len | Acc | Len | Acc | Len | Acc | Len | Acc | Len |
| Original | 93.8 | 1473 | 97.7 | 6675 | 66.3 | 8805 | 93.5 | 4324 | 87.8 | 5319 |
| ASC (calibrated ) | 93.6 | 1476 | 97.7 | 6715 | 66.5 | 8777 | 93.1 | 4349 | 87.7 | 5329 |
| Steer2Edit (best) | 90.4 | 1001 | 89.7 | 4045 | 42.4 | 5182 | 86.6 | 3322 | 77.3 | 3387 |
| Method | Avg Safety | Over-Refusal |
| Original | 49.9% | 60.1% |
| Steering ( =0.3) | 54.7% | 63.1% |
| Steering ( =0.5) | 60.5% | 65.8% |
| Steering ( =0.7) | 65.7% | 68.7% |
| Steering ( =1.0) | 75.7% | 72.6% |
| Steering ( =1.5) | 80.1% | 76.0% |
| Method | GCG | AdvLLM | Avg Safety | CSQA | GSM8K | CodeMMLU | Avg Utility |
| Original | 98.5 | 34.7 | 66.6 | 64.3 | 48.3 | 42.7 | 51.8 |
| Steering ( =0.5) | 99.3 | 49.3 | 74.3 | 62.4 | 45.9 | 43.1 | 50.4 |
| Steering ( =1.0) | 100.0 | 57.7 | 78.9 | 58.8 | 44.3 | 41.3 | 48.1 |
| Steering ( =1.5) | 93.7 | 65.0 | 79.4 | 46.3 | 37.6 | 29.8 | 37.9 |
| Steering ( =2.0) | 57.4 | 58.0 | 57.7 | 28.8 | 26.6 | 13.6 | 23.0 |
| Steer2Edit (best) | 99.9 | 94.1 | 97.0 | 44.8 | 26.9 | 35.5 | 35.7 |
| Method | Config | GCG | AdvLLM | Avg Safety | CSQA | GSM8K | CodeMMLU | Avg Utility |
| Original | – | 99.0 | 48.8 | 73.9 | 78.3 | 86.1 | 65.5 | 76.6 |
| Steering | =0.3 | 99.0 | 62.6 | 80.8 | 77.7 | 85.7 | 67.4 | 77.0 |
| Steering | =0.5 | 98.3 | 72.5 | 85.4 | 76.9 | 84.7 | 66.5 | 76.0 |
| Steering | =1.0 | 77.6 | 82.9 | 80.3 | 63.7 | 62.7 | 62.1 | 62.8 |
| Steering | =1.2 | 76.0 | 76.9 | 76.5 | 44.7 | 44.6 | 48.2 | 45.8 |
| Steer2Edit | =0.3, =0.5, =0.5 | 98.1 | 98.3 | 98.2 | 74.2 | 85.0 | 68.7 | 76.0 |
| Method | Config | GCG | AdvLLM | Avg Safety | CSQA | GSM8K | CodeMMLU | Avg Utility |
| Original | – | 100.0 | 3.0 | 51.5 | 36.2 | 15.2 | 50.0 | 33.8 |
| Steering | =0.18 | 100.0 | 44.0 | 72.0 | 11.8 | 2.1 | 0.0 | 4.6 |
| Steering | =0.2 | 100.0 | 59.0 | 79.5 | 2.0 | 1.8 | 0.0 | 1.3 |
| Steering | =0.22 | 62.0 | 22.0 | 42.0 | 0.0 | 0.3 | 0.0 | 0.1 |
| Steer2Edit | =0.3, =0.5, =0.9 | 100.0 | 76.0 | 88.0 | 19.2 | 0.6 | 36.6 | 18.8 |
| Method | Configuration | GCG | AdvLLM | Avg safety | CSQA | CodeMMLU | GSM8K | Avg utility |
| Original | – | 70.0 | 29.8 | 49.9 | 54.2 | 28.3 | 24.1 | 35.5 |
| ITI (highest safety) | 83.6 | 80.1 | 81.8 | 11.5 | 6.3 | 9.2 | 9.0 | |
| ITI (trade-off) | 81.6 | 59.4 | 70.5 | 37.1 | 23.2 | 21.2 | 27.2 | |
| AUSteer (highest safety) | Attention, | 81.9 | 93.0 | 87.5 | 25.2 | 16.1 | 6.6 | 16.0 |
| AUSteer (trade-off) | FFN, | 83.9 | 70.6 | 77.2 | 47.5 | 23.1 | 20.1 | 30.2 |
| Steer2Edit | – | 95.1 | 81.2 | 88.2 | 45.0 | 20.1 | 18.9 | 28.0 |
| Claude Opus 4.6 | |||
| QwQ-32B | TRUE | FALSE | Total |
| TRUE | 222 | 55 | 277 |
| FALSE | 11 | 120 | 131 |
| Total | 233 | 175 | 408 |
| Model / labeling judge | TruthfulQA | CSQA | CodeMMLU | GSM8K | Avg utility |
| Original | 40.3 | 71.8 | 51.0 | 74.7 | 65.8 |
| Edit, QwQ-32B | 53.1 | 62.7 | |||
| Edit, Claude Opus 4.6 | 50.6 | 66.3 |
Language Model Editing
Language model editing changes a model’s knowledge or behavior, or revises its outputs, while aiming to preserve unrelated capabilities.
See all 85 papers