mlsanityData to 14 Oct 2026

Beyond Linear Steering: Unified Multi-Attribute Control for Language Models

arXiv 2505.24535 ↗First posted 30 May 2025cs.LG cs.AI cs.CLcited by 25

Not on the map. Why?

Abstract (arXiv)

Controlling multiple behavioral attributes in large language models (LLMs) at inference time is a challenging problem due to interference between attributes and the limitations of linear steering methods, which assume additive behavior in activation space and require per-attribute tuning. We introduce K-Steering, a unified and flexible approach that trains a single non-linear multi-label classifier on hidden activations and computes intervention directions via gradients at inference time. This avoids linearity assumptions, removes the need for storing and tuning separate attribute vectors, and allows dynamic composition of behaviors without retraining. To evaluate our method, we propose two new benchmarks, ToneBank and DebateMix, targeting compositional behavioral control. Empirical results across 3 model families, validated by both activation-based classifiers and LLM-based judges, demonstrate that K-Steering outperforms strong baselines in accurately steering multiple behaviors.

Key facts. Cited by 25 papers in the map (data through 14 Oct 2026). Builds on 26 of its 43 references that are in the map. First posted on arXiv 30 May 2025.

Builds on

26 of 43 references.

References · 26

  1. 2022Training language models to follow instructions with human feedbackcited by 15,971
  2. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,720
  3. 2016Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddingscited by 1,312
  4. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,302
  5. 2023Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!cited by 1,154
  6. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  7. 2024Refusal in Language Models Is Mediated by a Single Directioncited by 1,032
  8. 2022Mass-Editing Memory in a Transformercited by 1,001
  9. 2019Plug and Play Language Models: A Simple Approach to Controlled Text Generationcited by 799
  10. 2023The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasetscited by 712
  11. 2023Safe RLHF: Safe Reinforcement Learning from Human Feedbackcited by 656
  12. 2023An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-tuningcited by 639
14 more
  1. 2022Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Taskcited by 431
  2. 2023RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedbackcited by 412
  3. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 218
  4. 2025Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMscited by 172
  5. 2023The Hydra Effect: Emergent Self-repair in Language Model Computationscited by 156
  6. 2024Collective Constitutional AI: Aligning a Language Model with Public Inputcited by 123
  7. 2019Attenuating Bias in Word Vectorscited by 64
  8. 2025A Unified Understanding and Evaluation of Steering Methodscited by 46
  9. 2025Multi-Attribute Steering of Language Models via Targeted Interventioncited by 46
  10. 2024Extending Activation Steering to Broad Skills and Multiple Behaviourscited by 40
  11. 2024Towards Reliable Evaluation of Behavior Steering Interventions in LLMscited by 36
  12. 2024Residual Stream Analysis with Multi-Layer SAEscited by 19
  13. 2025Effectively Steer LLM To Follow Preference via Building Confident Directionscited by 7
  14. 2023Parameter-Efficient Tuning Helps Language Model Alignmentcited by 4

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 3; 2026: 22.

25 papers.

Sort:

By area:

  1. 2025Angular Steering: Behavior Control via Rotation in Activation Spacecited by 53
  2. 2025Adaptive Multi-Subspace Representation Steering for Attribute Alignment in Large Language Modelscited by 16
  3. 2025PIXEL: Adaptive Steering Via Position-wise Injection with eXact Estimated Levels under Subspace Calibrationcited by 11
  4. 2026Curveball Steering: The Right Direction To Steer Isn't Always Linearcited by 11
  5. 2026Beyond Steering Vector: Flow-based Activation Steering for Inference-Time Interventioncited by 6
  6. 2026Global Evolutionary Steering: Refining Activation Steering Control via Cross-Layer Consistencycited by 5
  7. 2026The Cylindrical Representation Hypothesis for Language Model Steeringcited by 5
  8. 2026Steer Like the LLM: Activation Steering that Mimics Promptingcited by 5
  9. 2026ORBIT: Training-Free Multi-Attribute Behavioral Steering via Orthogonal Subspace Rotationcited by 4
  10. 2026The Lattice Representation Hypothesis of Large Language Modelscited by 3
  11. 2026Steering Without Breaking: Mechanistically Informed Interventions for Discrete Diffusion Language Modelscited by 3
  12. 2026Deployable Per-Instance Multi-Layer Activation Steering for Large Language Modelscited by 3
13 more
  1. 2026On The Effectiveness-Fluency Trade-Off In LLM Conditioning: A Systematic Studycited by 1
  2. 2026DreamReader: An Interpretability Toolkit for Text-to-Image Modelscited by 0
  3. 2026GEMS: Geometric Constraints Enable Multi-Semantic Superposition in LLMscited by 0
  4. 2026Policy Gradient Steering: Interventions from Behavioral Objectivescited by 0
  5. 2026The Illusion of Control: Why Bare Classifier Inversion Silently Fails in Concept-Bottleneck Text Generationcited by 0
  6. 2026Attribute-Based Activation Steering of LLMs for Group-Specific Explanation Generationcited by 0
  7. 2026Can Activation Steering Capture Multidimensional Authorship Style?cited by 0
  8. 2026Steering Interference Reflects the Model's Defaults, Not the Behavior Directionscited by 0
  9. 2026Adaptive Multi-Value Control in LLMs via Causal Activation Steeringcited by 0
  10. 2026How to Tame a Multi-Headed Hydra? Adaptive Multi-Category Safety Steering for Large Language Modelscited by 0
  11. 2026Kernelized Activation Steeringcited by 0
  12. 2026Hesitation Has a Geometry: Entropy-Trained Hyperbolic Probes for Sparse Activation Steeringcited by 0
  13. 2026First-Order Steering: Translating Weight Adaptation into Activation Steeringcited by 0