mlsanityData to 7 Oct 2026

Where Steering Signals Come From: Activation Source Selection in Activation Steering

arXiv 2607.25270 ↗First posted 28 Jul 2026cs.CL cs.AI cs.LGcited by 2

Not on the map. Why?

Abstract (arXiv)

Activation steering controls language models by adding vectors or features to hidden states at inference time, but the upstream source of these steering signals is often treated as a secondary detail. We study this source choice as activation source selection: the combination of source context and activation readout policy used to collect the hidden states from which a steering signal is built. Holding the downstream intervention fixed, we show across three instruction-tuned models and four steering task families that changing only the source activations substantially changes steering success. We further find that effective steering is not explained simply by whether the desired behavior appears in the source text. Instead, strong signals come from execution-boundary states, where the model is about to produce or continue the target behavior. This pre-/post-realization distinction explains why answer-based sources sometimes work: their useful component aligns with execution-boundary directions rather than target appearance alone. Building on this view, we introduce tail subtraction, which removes shared prompt and continuation semantics from boundary states and yields cleaner, more stable steering signals. Overall, our results suggest that steering depends on representations of what the model is about to do, not merely on what has already appeared.

Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Builds on 29 of its 34 references that are in the map. First posted on arXiv 28 Jul 2026.

Builds on

29 of 34 references.

References · 29

  1. 2024The Llama 3 Herd of Modelscited by 14,192
  2. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,928
  3. 2022Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?cited by 1,413
  4. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,359
  5. 2016Understanding intermediate layers using linear classifier probescited by 1,279
  6. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  7. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
  8. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
  9. 2023Steering Language Models With Activation Engineeringcited by 824
  10. 2022What Can Transformers Learn In-Context? A Case Study of Simple Function Classescited by 700
  11. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 381
  12. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 303
17 more
  1. 2023In-Context Learning Creates Task Vectorscited by 302
  2. 2018Analysis Methods in Neural Language Processing: A Surveycited by 276
  3. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 199
  4. 2024Programming Refusal with Conditional Activation Steeringcited by 191
  5. 2024Improving Instruction-Following in Language Models through Activation Steeringcited by 149
  6. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 133
  7. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 86
  8. 2023Improving Activation Steering in Language Models with Mean-Centringcited by 64
  9. 2024Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectorscited by 61
  10. 2025FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steeringcited by 37
  11. 2025Refusal Direction is Universal Across Safety-Aligned Languagescited by 37
  12. 2024Activation Scaling for Steering and Interpreting Language Modelscited by 33
  13. 2024Steering Large Language Models using Conceptors: Improving Addition-Based Activation Engineeringcited by 33
  14. 2025SAIF: A Sparse Autoencoder Framework for Interpreting and Steering Instruction Following of Language Modelscited by 32
  15. 2026Spherical Steering: Geometry-Aware Activation Rotation for Language Modelscited by 17
  16. 2025SAKE: Steering Activations for Knowledge Editingcited by 12
  17. 2017Contrastive Principal Component Analysiscited by 7

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026What Does Activation Steering Control? Attribution Across Answer Encodings and Output-Sensitive Subspacescited by 1
  2. 2026SteerCheck: Attribution Specificity and Alignment Leakage in Activation-Steering Auditscited by 1