Research map / Large language models / Language, safety and interpretability

Steer2Edit: From Activation Steering to Component-Level Editing

NeurIPS 2026 · Large language models › Language, safety and interpretability · approach: Task vectors and steering · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Steer2Edit converts mean-difference steering vectors into closed-form, training-free rank-1 weight edits on individual attention-head o_proj and MLP down_proj columns, improving attribute-utility trade-offs for safety, truthfulness, and reasoning-length control.

Problem

Activation steering applies a global, inference-time modification to hidden states, which often causes unfavorable attribute-utility trade-offs under strong control and breaks the standard forward pass assumed by optimized inference pipelines.

Approach

Treat steering vectors as diagnostic signals: for each attention head or MLP neuron, build a rank-1 weight edit with output direction set to the normalized steering vector, input direction set to W^T v, and magnitude given by a soft-thresholded, cosine-based component importance score under an Elastic-Net budget.

Key results

• At matched downstream performance, Steer2Edit improves safety (refusal rate) by up to 17.2%.
• Steer2Edit increases truthfulness by 9.8% at matched utility.
• Steer2Edit reduces reasoning length by 12.2% on average while maintaining comparable accuracy.

Surprising findings

• Unnormalized importance score (g_dot) variant collapses performance (overall 0.039), indicating raw activation magnitude across layers dominates edits.
• Truthfulness edits on Gemma-2-2B-IT are dominated by negative coefficients, suggesting gains come from suppressing hallucination-promoting components.

Limitations

• Relies on mean-difference steering vectors and a hyperparameter grid search on validation sets with per-model ranges.
• Evaluated on only two models per behavior setting and a limited set of benchmarks.

Experimental setup

Datasets: ADVBench Alpaca TruthfulQA GSM8K

Benchmarks: ADVBench GSM8K CodeMMLU CommonsenseQA TruthfulQA MATH-500 GPQA

Metrics: refusal rate truthful preference accuracy accuracy reasoning tokens

Baselines: activation steering (mean-difference) ThinkEdit DRefA full fine-tuning rank-1 LoRA

Builds on: representation engineering activation steering ROME ThinkEdit

Closest papers by meaning

• Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior (NeurIPS 2026 · similarity 0.89)
• The Sharp Directions Are Against You: Curvature Analysis of Activation Steering (NeurIPS 2026 · similarity 0.89)
• DISCO: Disentangled Communication Steering for Large Language Models (NeurIPS 2025 · similarity 0.88)
• Improving Instruction-Following in Language Models through Activation Steering (ICLR 2025 · similarity 0.88)
• Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs (NeurIPS 2026 · similarity 0.88)
• When Is Rank-1 Steering Cheap? Geometry, Granularity, and Budgeted Search (NeurIPS 2026 · similarity 0.87)