Disentangling Steering Vectors

• URL: http://arxiv.org/abs/2609.07037v1
• Date: Mon, 07 Sep 2026 04:45:32 GMT
• Title: Disentangling Steering Vectors
• Abstract summary: We propose a framework to explicitly isolate individual and semantically consistent features from composite directions.<n>Specifically, we pair positive and negative activations and take their differences to generate a set of instance-level steering vectors.<n>We show that our method yields a set of semantically consistent basis vectors whose steering effects are mutually distinguishable.
• Score: 13.671243729016568
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Activation steering has emerged as a lightweight, inference-time approach to control the behavior of Large Language Models (LLMs). However, traditional steering vectors used to intervene in LLMs' activations, such as those derived from the difference-in-means method, tend to entangle multiple semantic and stylistic concepts into a single composite direction, leading to unpredictable steering effects. Our core objective is to disentangle this composite direction into its constituent concepts. To this end, we propose Steering Vector Dissection, a framework to explicitly isolate individual and semantically consistent features from these composite directions. Specifically, we pair positive and negative activations and take their differences to generate a set of instance-level steering vectors, and train a dedicated Sparse Autoencoder (SAE) directly on them. Quantitative evaluations across two datasets, two models, and two intervention depths show that our method yields a set of semantically consistent basis vectors whose steering effects are mutually distinguishable. Furthermore, we show that this disentanglement enables precise control over model behaviors.

Related papers

• Where Steering Signals Come From: Activation Source Selection in Activation Steering [71.96112659293725]
  We find that effective steering is not explained simply by whether the desired behavior appears in the source text.<n>Strong signals come from execution-boundary states, where the model is about to produce or continue the target behavior.<n>Building on this view, we introduce tail subtraction, which removes shared prompt and continuation semantics from boundary states.
  arXiv  Detail & Related papers  (2026-07-28T04:18:24Z)
• On the Limits of Steering Vectors for Preference-Aligned Generation [12.350800078736128]
  We study the limits of steering vector generalization along three dimensions: trait expressibility, task transfer, and multi-trait composition.<n>Our results suggest that steering vectors face meaningful limits as a general-purpose tool for preference alignment.
  arXiv  Detail & Related papers  (2026-07-02T07:18:36Z)
• What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal [53.189667624047416]
  We show that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer.<n>We show that steering vectors can be sparsified by up to 90-99% while retaining most performance.
  arXiv  Detail & Related papers  (2026-04-09T17:57:14Z)
• ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment [49.68063561145927]
  We propose a unified ordinary differential equations (ODEs)-based theoretical framework for activation steering.<n>We introduce ODESteer, a kind of ODE-based steering guided by barrier functions.<n>Compared to state-of-the-art activation steering methods, ODESteer achieves consistent empirical improvements.
  arXiv  Detail & Related papers  (2026-02-19T17:13:44Z)
• Concept Heterogeneity-aware Representation Steering [6.247724933679344]
  Representation steering offers a mechanism for controlling the behavior of large language models (LLMs) by intervening on internal activations at inference time.<n>Most existing methods rely on a single global steering direction, typically obtained via difference-in-means over contrastive datasets.<n>In this work, we view representation steering through the lens of optimal transport (OT), noting that standard difference-in-means steering implicitly corresponds to the OT map between two unimodal Gaussian distributions.
  arXiv  Detail & Related papers  (2026-02-13T11:07:23Z)
• Why Steering Works: Toward a Unified View of Language Model Parameter Dynamics [81.80010043113445]
  Local weight fine-tuning, LoRA-based adaptation, and activation-based interventions are studied in isolation.<n>We present a unified view that frames these interventions as dynamic weight updates induced by a control signal.<n>Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility.
  arXiv  Detail & Related papers  (2026-02-02T17:04:36Z)
• Steering Risk Preferences in Large Language Models by Aligning Behavioral and Neural Representations [4.029252551781513]
  We propose a principled approach for uncovering steering vectors.<n>We focus on extracting latent risk preferences from large language models.<n>We show that the resulting steering vectors successfully and reliably modulate LLM outputs in line with the targeted behavior.
  arXiv  Detail & Related papers  (2025-05-16T18:23:10Z)
• Activation Scaling for Steering and Interpreting Language Models [55.59689963561315]
  We argue that successfully intervening on a model is a prerequisite for interpreting its internal workings. We establish a three-term objective: a successful intervention should flip the correct with the wrong token and vice versa. Using gradient-based optimization, this objective lets us learn (and later evaluate) a specific kind of efficient and interpretable intervention.
  arXiv  Detail & Related papers  (2024-10-07T12:01:32Z)
• Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization [34.05163996072159]
  "steering vectors" are extracted from the activations of human preference data. This work proposes an innovative approach that could produce more effective steering vectors through bi-directional preference optimization. Our method is designed to allow steering vectors to directly influence the generation probability of contrastive human preference data pairs.
  arXiv  Detail & Related papers  (2024-05-28T05:10:40Z)
• Steering Llama 2 via Contrastive Activation Addition [41.54815073311959]
  Contrastive Activation Addition (CAA) is a method for steering language models by modifying their activations during forward passes. CAA accurately steers model outputs and sheds light on how high-level concepts are represented in Large Language Models (LLMs)
  arXiv  Detail & Related papers  (2023-12-09T04:40:46Z)
• On the Role of Bidirectionality in Language Model Pre-Training [85.14614350372004]
  We study the role of bidirectionality in next token prediction, text infilling, zero-shot priming and fine-tuning. We train models with up to 6.7B parameters, and find differences to remain consistent at scale.
  arXiv  Detail & Related papers  (2022-05-24T02:25:05Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.