mlsanityData to 7 Oct 2026

What Does Activation Steering Control? Attribution Across Answer Encodings and Output-Sensitive Subspaces

arXiv 2608.22985 ↗First posted 24 Aug 2026cs.CLcited by 1

Not on the map. Why?

Abstract (arXiv)

Activation steering is often evaluated under the answer encoding used to construct the direction. A reported gain may reflect the intended judgment or compatibility with answer identifiers seen during construction. We introduce Cross-Encoding Steering Evaluation, which freezes an intervention while re-encoding answers to the same held-out items. On NormBank, after A/B/C identifiers are reassigned, contrastive activation addition (CAA) induces larger target-versus-source score changes for the extraction indices than for the semantic labels under the new mapping. We call this extraction-index following. Varying identifier vocabulary (A/B/C, X/Y/Z, or 1/2/3) and row order shows that the effect tracks extraction index rather than row position. After matching direction norms across layers, extraction-index following emerges mainly at later depths. A low-rank output-sensitive component containing 15.4% of the direction's squared norm retains 96.3% of this effect. An Inference-Time Intervention (ITI)-style method also favors extraction-index over semantic-label following on NormBank in three models. In aggregate, MNLI favors extraction-index following, whereas Social Chemistry 101 (SC101) favors semantic-label following. Multiple-choice and open-ended evaluations can yield different behavioral conclusions. Thus, a steering gain under one answer encoding does not by itself identify what the intervention controls.

Key facts. Cited by 1 paper in the map (data through 7 Oct 2026). Builds on 28 of its 28 references that are in the map. First posted on arXiv 24 Aug 2026.

Builds on

28 of 28 references.

References · 28

  1. 2024The Llama 3 Herd of Modelscited by 14,192
  2. 2024Qwen2.5 Technical Reportcited by 4,713
  3. 2023Mistral 7Bcited by 3,065
  4. 2017A Broad-Coverage Challenge Corpus for Sentence Understanding through Inferencecited by 2,665
  5. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,928
  6. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  7. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
  8. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
  9. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 679
  10. 2023Eliciting Latent Predictions from Transformers with the Tuned Lenscited by 579
  11. 2023Large Language Models Are Not Robust Multiple Choice Selectorscited by 454
  12. 2020Social Chemistry 101: Learning to Reason about Social and Moral Normscited by 253
16 more
  1. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 130
  2. 2024AtP*: An efficient and scalable method for localizing LLM behaviour to componentscited by 100
  3. 2023NormBank: A Knowledge Bank of Situational Social Normscited by 60
  4. 2024Towards Reliable Evaluation of Behavior Steering Interventions in LLMscited by 33
  5. 2023Instruction-following Evaluation through Verbalizer Manipulationcited by 30
  6. 2024Exploring and steering the moral compass of Large Language Modelscited by 15
  7. 2025Internal Value Alignment in Large Language Models through Controlled Value Vector Activationcited by 14
  8. 2026On the Non-Identifiability of Steering Vectors in Large Language Modelscited by 8
  9. 2026Where Do Apparent LLM Clinical Triage Failures Arise? Localizing the Multiple-Choice Format Effectcited by 4
  10. 2026Controllable Value Alignment in Large Language Models through Neuron-Level Editingcited by 3
  11. 2026How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularitiescited by 3
  12. 2026On the Context Sensitivity of LLM Moral Judgmentcited by 3
  13. 2026Cultural Value Alignment Via Latent Activation Steering in Large Language Modelscited by 3
  14. 2026Scenario-based Probing and Steering Cultural Values in Large Language Models--Extended Versioncited by 2
  15. 2026Where Steering Signals Come From: Activation Source Selection in Activation Steeringcited by 2
  16. 2026DecodeShare: Tracing the Shared Subspace of LLM Decode-Time Decisionscited by 1

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026Replay-Gated Neural Execution: Decoupling Persistent Behavioral Specifications from Neural Realizations in Frozen Language Modelscited by 0