mlsanityData to 7 Oct 2026

Probabilistic Concept-Aware Steering for Trustworthy LLM Inference

arXiv 2607.18259 ↗First posted 15 May 2026cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

Steering vectors (SVs), an inference-time intervention technique for large language models (LLMs), guide the generation process by adding a concept-specific direction vector to intermediate activations during inference. However, existing SV methods frequently yield representation-incoherent behaviors that undermine interpretability and fine-grained control, largely because prior work has focused on binary positive-negative steering evaluation while employing discrete clustering metrics that fail to capture the continuous spectrum of semantic alignment. In this work, we present the Probabilistic Concept-Aware Steering (PCS) framework for LLM inference. PCS preserves original task competence while providing controllable, safety-oriented semantic bias through concept-driven steering-vector retrieval and probabilistic strength calibration.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Builds on 35 of its 45 references that are in the map. First posted on arXiv 15 May 2026.

Builds on

35 of 45 references.

References · 35

  1. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  2. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,009
  3. 2023Steering Language Models With Activation Engineeringcited by 824
  4. 2022Discovering Language Model Behaviors with Model-Written Evaluationscited by 805
  5. 2023The Linear Representation Hypothesis and the Geometry of Large Language Modelscited by 679
  6. 2023GPT4RoI: Instruction Tuning Large Language Model on Region-of-Interestcited by 300
  7. 2023How Abilities in Large Language Models are Affected by Supervised Fine-tuning Data Compositioncited by 222
  8. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 199
  9. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 130
  10. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimizationcited by 112
  11. 2023Batch Calibration: Rethinking Calibration for In-Context Learning and Prompt Engineeringcited by 91
  12. 2024Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categoriescited by 74
23 more
  1. 2024Semantics-Adaptive Activation Intervention for LLMs via Dynamic Steering Vectorscited by 61
  2. 2024Understanding LLMs: A Comprehensive Overview from Training to Inferencecited by 58
  3. 2024Style Vectors for Steering Generative Large Language Modelcited by 56
  4. 2023Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuningcited by 45
  5. 2024Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMscited by 39
  6. 2024ControlMLLM: Training-Free Visual Prompt Learning for Multimodal Large Language Modelscited by 33
  7. 2024Steering Large Language Models using Conceptors: Improving Addition-Based Activation Engineeringcited by 33
  8. 2024Towards Reliable Evaluation of Behavior Steering Interventions in LLMscited by 33
  9. 2025SafeSwitch: Steering Unsafe LLM Behavior via Internal Activation Signalscited by 29
  10. 2025Steering off Course: Reliability Challenges in Steering Language Modelscited by 29
  11. 2024SemScore: Automated Evaluation of Instruction-Tuned LLMs based on Semantic Textual Similaritycited by 18
  12. 2024Contextualization Distillation from Large Language Model for Knowledge Graph Completioncited by 18
  13. 2025Improved Representation Steering for Language Modelscited by 17
  14. 2025Differentially Private Steering for Large Language Model Alignmentcited by 15
  15. 2025Current and Future Use of Large Language Models for Knowledge Workcited by 10
  16. 2025Evolving Prompts In-Context: An Open-ended, Self-replicating Perspectivecited by 8
  17. 2026COLD-Steer: Steering Large Language Models via In-Context One-step Learning Dynamicscited by 7
  18. 2022Conceptor-Aided Debiasing of Large Language Modelscited by 4
  19. 2025W2S-AlignTree: Weak-to-Strong Inference-Time Alignment for Large Language Models via Monte Carlo Tree Searchcited by 4
  20. 2026Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejectioncited by 4
  21. 2024Enhancing Fast Feed Forward Networks with Load Balancing and a Master Leaf Nodecited by 1
  22. 2025EnSToM: Enhancing Dialogue Systems with Entropy-Scaled Steering Vectors for Topic Maintenancecited by 1
  23. 2025Explicit Vulnerability Generation with LLMs: An Investigation Beyond Adversarial Attackscited by 1

Built on by

Not cited yet.