mlsanityData to 14 Oct 2026

Steered Generation via Gradient-Based Optimization on Sparse Query Features

arXiv 2605.23040 ↗First posted 21 May 2026cs.LGcited by 0

On the map: Mechanistic interpretability → Activation steering methods

Abstract (arXiv)

Latent steering exploits internal representations of Large Language Models (LLMs) to guide generation, yet interventions on dense states can entangle distinct semantic features. In this paper, we investigate attention query activations as a high-fidelity site for precise control, hypothesizing that manipulating the attention mechanism itself offers sharper steerability than general state interventions. We introduce Prototype-Based Sparse Steering, a framework that applies Sparse Autoencoders (SAEs) specifically to query activations, to decompose them into interpretable features, then apply gradient-based optimization during inference to align the sparse representation with class prototypes of target behaviors. To validate this architectural insight, we first analyze the mechanism in Textualized Gridworld, a controlled environment for verifiable planning constraints. We demonstrate that optimizing sparse query features enables effective navigation of rigid planning requirements (i.e., safe vs. short paths), confirming the method's ability to satisfy objective rules. We then demonstrate the framework's versatility by training SAEs on a high-dimensional educational domain, where the framework steers the cognitive complexity of feedback (i.e., Bloom's Taxonomy). Our experiments establish that sparse query representations provide the necessary disentanglement for unified, interpretable control over both logical planning and stylistic nuance.

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 85 of its 106 references that are in the map. First posted on arXiv 21 May 2026.

Builds on

85 of 106 references.

Closest ancestors · 10

Same direction, newest first

  1. 2025DISCO: Disentangled Communication Steering for Large Language Models20 Sep
  2. 2025LinEAS: End-to-end Learning of Activation Steering with a Distributional Loss11 Mar
  3. 2025Personalized Text Generation with Contrastive Activation Steering7 Mar
  4. 2025Steering Large Language Model Activations in Sparse Spaces28 Feb
  5. 2025A Unified Understanding and Evaluation of Steering Methods4 Feb
  6. 2025Interpretable Steering of Large Language Models with Feature Guided Activation Additions17 Jan
  7. 2024Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectors16 Oct
  8. 2024Improving Instruction-Following in Language Models through Activation Steering15 Oct
  9. 2023Steering Llama 2 via Contrastive Activation Addition9 Dec
  10. 2022Extracting Latent Steering Vectors from Pretrained Language Models10 May

Other references · 75

  1. 2020Language Models are Few-Shot Learnerscited by 32,419
  2. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,459
  3. 2023Lost in the Middle: How Language Models Use Long Contextscited by 3,455
  4. 2017Prototypical Networks for Few-shot Learningcited by 3,015
  5. 2021TruthfulQA: Measuring How Models Mimic Human Falsehoodscited by 2,975
  6. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  7. 2024Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phonecited by 1,917
  8. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  9. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  10. 2023Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!cited by 1,154
  11. 2022Editing Models with Task Arithmeticcited by 1,120
  12. 2022Toy Models of Superpositioncited by 971
38 more
  1. 2023Progress measures for grokking via mechanistic interpretabilitycited by 839
  2. 2019Plug and Play Language Models: A Simple Approach to Controlled Text Generationcited by 799
  3. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 716
  4. 2015Sequence Level Training with Recurrent Neural Networkscited by 576
  5. 2024Scaling and evaluating sparse autoencoderscited by 553
  6. 2021DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Expertscited by 441
  7. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  8. 2015Autoencoding beyond pixels using a learned similarity metriccited by 407
  9. 2024Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoderscited by 300
  10. 2024Massive Activations in Large Language Modelscited by 294
  11. 2024Long-context LLMs Struggle with Long In-context Learningcited by 286
  12. 2020Understanding the Role of Individual Units in a Deep Neural Networkcited by 255
  13. 2023In-context Vectors: Making In Context Learning More Effective and Controllable Through Latent Space Steeringcited by 245
  14. 2019Analyzing the Structure of Attention in a Transformer Language Modelcited by 243
  15. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 218
  16. 2013k-Sparse Autoencoderscited by 215
  17. 2024Improving Dictionary Learning with Gated Sparse Autoencoderscited by 179
  18. 2020Compositional Explanations of Neuronscited by 172
  19. 2022The Unreliability of Explanations in Few-shot Prompting for Textual Reasoningcited by 169
  20. 2019Adaptively Sparse Transformerscited by 155
  21. 2023Inspecting and Editing Knowledge Representations in Language Modelscited by 141
  22. 2022A Survey of Controllable Text Generation using Transformer-based Pre-trained Language Modelscited by 129
  23. 2020A Distributional Approach to Controlled Text Generationcited by 119
  24. 2021Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factorscited by 108
  25. 2024The Butterfly Effect of Altering Prompts: How Small Changes and Jailbreaks Affect Large Language Model Performancecited by 100
  26. 2024Steering Language Model Refusal with Sparse Autoencoderscited by 95
  27. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 92
  28. 2023Tell Your Model Where to Attend: Post-hoc Attention Steering for LLMscited by 91
  29. 2025SAEs Are Good for Steering -- If You Select the Right Featurescited by 78
  30. 2022Polysemanticity and Capacity in Neural Networkscited by 77
  31. 2023Word Embeddings Are Steers for Language Modelscited by 70
  32. 2022Mix and Match: Learning-free Controllable Text Generation using Energy Language Modelscited by 66
  33. 2021Prompt Waywardness: The Curious Case of Discretized Interpretation of Continuous Promptscited by 64
  34. 2024Steering Knowledge Selection Behaviours in LLMs via SAE-Based Representation Engineeringcited by 63
  35. 2025Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignmentcited by 62
  36. 2025Multi-Attribute Steering of Language Models via Targeted Interventioncited by 46
  37. 2025I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoderscited by 45
  38. 2025FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steeringcited by 40

50 most cited of 75 All 75 (JSON)

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2025Steered Generation via Gradient Descent on Sparse Features
  2. 2026A Coin Flip Per Token: Bernoulli Sparse Steering of Large Language Models
  3. 2026MetaSteer: Context-Conditioned, nonlinear Steering via Attention-Projection Adaptation
  4. 2025Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atoms
  5. 2026Minimally Invasive Steering of Language Models
  6. 2024Steering Large Language Models using Conceptors: Improving Addition-Based Activation Engineering
  7. 2026Don't Lose Focus: Activation Steering via Key-Orthogonal Projections
  8. 2026Control Reinforcement Learning: Interpretable Token-Level Steering of LLMs via Sparse Autoencoder Features