mlsanityData to 14 Oct 2026

Steering by Influence: Curvature Aware Data Weighting for Activation Steering

arXiv 2610.06383 ↗First posted 5 Oct 2026cs.LG cs.AI cs.CLcited by 0

On the map: Mechanistic interpretability → Activation steering methods

Abstract (arXiv)

Inference-time steering offers cheap, fine-grained control over a language model's outputs by estimating a concept's representation in activation space and shifting activations towards it. Existing methods build these representations from activation averages over contrastive datasets. These averages incorporate unrelated concepts and noise, and are dominated by a few tokens, meaning the activation transport encodes token-level rather than thematic concepts. In this work, we steer towards examples that most express a concept thematically, rather than towards an expectation over all. We identify these examples using influence functions, which estimate how much each data point contributes to a model's representation of a concept. Unlike simple model activation similarity, they incorporate the curvature of the model's loss landscape, allowing them to capture concept-relevant relationships beyond superficial token-level similarity. We then propose influence-weighted activation transport, which uses optimal transport to steer activations of non-concept text towards those of concept text, weighting concept examples by their influence scores. We evaluate on toxicity suppression (Jigsaw), object-based concept induction (OneSec) and truthfulness induction (TruthfulQA), outperforming existing activation-transport baselines. We track capability after steering using perplexity and MMLU accuracy, finding that our method improves steering while largely preserving model quality. We further show that influence functions capture concept-relevant information that activation-based methods miss with the two approaches ranking data points significantly differently. Together, these results demonstrate the value of curvature-aware influence information for activation steering.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Activation steering methods, one of 13 directions in Mechanistic interpretability. Builds on 23 of its 62 references that are in the map. First posted on arXiv 5 Oct 2026.

Builds on

23 of 62 references.

Closest ancestors · 3

Same direction, newest first

  1. 2026Concept Heterogeneity-aware Representation Steering13 Feb
  2. 2023Improving Activation Steering in Language Models with Mean-Centring6 Dec
  3. 2023Steering Language Models With Activation Engineering20 Aug

Other references · 20

  1. 2024The Llama 3 Herd of Modelscited by 14,715
  2. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,380
  3. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,372
  4. 2020Scaling Laws for Neural Language Modelscited by 6,586
  5. 2024Gemma 2: Improving Open Language Models at a Practical Sizecited by 1,998
  6. 2022GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformerscited by 1,919
  7. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,336
  8. 2022Toy Models of Superpositioncited by 971
  9. 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 767
  10. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
  11. 2023Studying Large Language Model Generalization with Influence Functionscited by 326
  12. 2020Influence Functions in Deep Learning Are Fragilecited by 252
8 more
  1. 2024The rising costs of training frontier AI modelscited by 95
  2. 2024Representation Surgery: Theory and Practice of Affine Steeringcited by 50
  3. 2024On Effects of Steering Latent Representation for Large Language Model Unlearningcited by 32
  4. 2023ASDL: A Unified Interface for Gradient Preconditioning in PyTorchcited by 21
  5. 2024Delta-Influence: Unlearning Poisons via Influence Functionscited by 13
  6. 2025The Blessing and Curse of Dimensionality in Safety Alignmentcited by 10
  7. 2026Infusion: Shaping Model Behavior by Editing Training Data via Influence Functionscited by 5
  8. 2026Detecting Contaminated Code-Generation Prompt Batches via Influence Functionscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Disentangling Steering Vectors
  2. 2026Genesis Steering: Better Activation Steering Through Training Data
  3. 2024Activation Scaling for Steering and Interpreting Language Models
  4. 2026Contrastive Covariance Subspaces for Activation Steering
  5. 2025Mitigating Content Effects on Reasoning in Language Models through Fine-Grained Activation Steering
  6. 2026Beyond Steering Vector: Flow-based Activation Steering for Inference-Time Intervention
  7. 2026COLD-Steer: Steering Large Language Models via In-Context One-step Learning Dynamics
  8. 2026Signatures of Steerability in Activation Space of Language Models