Abstract

Steering methods influence Large Language Model behavior by identifying semantic directions in hidden representations, and 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.

To alleviate the trade-offs, 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, Steer2Edit selectively redistributes behavioral influence across individual attention heads and MLP neurons, yielding interpretable edits that preserve the standard forward pass.

Figure 1. Overview. Steer2Edit converts the steering signal into component-level rank-1 weight edits \(\Delta W_i = \lambda_i u_i k_i^\top\). For each component, the edit is constructed by aligning the output direction, choosing an input direction that triggers the edit only on relevant inputs, and allocating magnitude under a global budget.

Method

Steer2Edit is a principled framework for component-level weight editing based on given steering vectors. We parameterize each edit as a rank-1 update and derive its form by decomposing the problem into three parts: [Step 1] identifying the output-space direction that preserves semantic invariance, [Step 2] the input-space direction that aligns the edit with the component’s intrinsic semantic contribution, and [Step 3] the scalar magnitude that allocates edit strength under a global regularization budget.

Assumption and Setting

For each editable component \(W_i \in \mathbb{R}^{d_{\text{out}} \times d_{\text{in}}}\), we assume the existence of a steering vector \(v_i \in \mathbb{R}^{d_{\text{out}}}\) extracted from the same representation space into which \(W_i\) writes. Our goal is to modify each component \(W_i\) so that the resulting update \(\Delta W_i\) alters the model’s behavior along the semantic direction represented by \(v_i\). We parameterize the edit as a rank-1 update:

where \(u_i\) is an output-space direction, \(k_i\) is an input-space direction, and \(\lambda_i\) is a scalar magnitude that need to be determined.

Unified Weight Editing Rule

Putting Steps 1–3 together, Steer2Edit yields a single closed-form update per component that is directionally selective, input-selective, and budget-aware.

Each editable component \(W_i\) receives the rank-1 update:

where \(g_i = \cos(v_i,\, W_i \mu_i)\) is the component importance score.

• Directionally selective: modifies only the projection along the semantic direction \(v_i\).
• Input-selective: triggers in proportion to the component’s intrinsic semantic behavior \(W_i^\top v_i\).
• Budget-aware: magnitude \(\lambda_i\) is allocated by an Elastic-Net regularizer under budget \(\rho\) and sparsity \(\alpha\).

Experiments: Attribute-Utility Trade-offs

While our goal is to control specific model behaviors through Steer2Edit updates, it is equally important to ensure that these interventions do not degrade the model’s original utility. Prior work on standard activation steering typically reveals a clear trade-off: as the target attribute is strengthened, the model’s general capabilities often deteriorate. This makes it crucial to demonstrate that Steer2Edit achieves a more favorable trade-off between behavioral control and overall performance.

We compare Steer2Edit with standard activation steering across three behavior control settings. In each case, the objective is to improve the target attribute while preserving performance on unrelated tasks, such as mathematical reasoning and coding.

Setting 1: Safety Alignment

1. Trade-off Analysis

Ideally, we want to be in the top-right corner (High Safety, High Utility). The gray lines show that standard steering sacrifices utility to gain safety. Steer2Edit (Red Stars) breaks this frontier, achieving significantly higher refusal rates while maintaining high utility.

Figure 2. Safety-utility trade-off on LLaMA-2-7B and Mistral-7B. Steer2Edit dominates the trade-off frontier.

2. Edited Component Analysis

This heatmap plots the best edit magnitude \(\lambda\) for every attention head in the model.
Red (Positive): The edit reinforces this head (darker means larger).
Blue (Negative): The edit suppresses this head (darker means larger).
Finding: Safety is managed by a highly sparse set of attention heads concentrated in the later layers.

Figure 3. Component edit distribution. Safety alignments are controlled with very few attention heads.

Setting 2: Truthfulness

1. Trade-off Analysis

We aim for the top-right (High Truthfulness, High Utility). Strong activation steering (gray lines) improves truthfulness but causes utility to plummet. Steer2Edit maintains stable utility even as truthfulness increases.

Figure 4. Truthfulness-utility trade-off on Gemma2-2B and LLaMA3-8B.

2. Edited Component Analysis

The heatmap shows edit coefficients across layers.
Finding: Truthfulness edits are also sparse but distributed across both early and late layers. Notably, many coefficients are Blue (Negative) for Gemma2-2B, suggesting this model's truthfulness is highly related to attention heads that promote hallucinations.

Figure 5. Truthfulness control are attention-dominated.

Setting 3: Efficient Reasoning

1. Trade-off Analysis

We want lower reasoning length (y-axis down) with high accuracy (x-axis right). Standard steering reduces length but destroys accuracy. Steer2Edit successfully shortens traces while keeping accuracy high.

Figure 6. Accuracy-efficiency trade-off on Qwen3-4B-Thnking and Nemotron-7B.

2. Edited Component Analysis

Finding: In sharp contrast to safety, reasoning length is mostly controlled by MLP Neurons (not attention heads). The edits are dense and distributed across the network, indicating efficiency is a broad, global behavior spanning many neurons.

Figure 7. Reasoning Efficiency is governed by dense, distributed MLP neurons.

BibTeX

@article{sun2026steer2edit,
  title={Steer2Edit: From Activation Steering to Component-Level Editing},
  author={Sun, Chung-En and Yan, Ge and Wang, Zimo and Weng, Tsui-Wei},
  journal={arXiv preprint arXiv:2602.09870},
  year={2026}
}