mlsanityData to 14 Oct 2026

Beyond Prompt Engineering: Robust Behavior Control in LLMs via Steering Target Atoms

arXiv 2505.20322 ↗First posted 23 May 2025cs.CL cs.AI cs.CVcited by 24

Not on the map. Why?

Abstract (arXiv)

Precise control over language model generation is vital for ensuring both safety and reliability. Although prompt engineering and steering are commonly used to intervene in model behaviors, the vast number of parameters in models often results in highly intertwined internal representations. This interdependency can limit control precision and sometimes lead to unintended side effects. Recent research has explored the use of sparse autoencoders (SAE) to disentangle knowledge in high-dimensional spaces for steering. However, these applications have been limited to toy tasks owing to the nontrivial issue of locating atomic knowledge components. In this paper, we propose Steering Target Atoms (STA), a novel method that isolates and manipulates disentangled knowledge components to enhance safety. Comprehensive experiments demonstrate the effectiveness of our approach. Further analysis reveals that steering exhibits superior robustness and flexibility, particularly in adversarial scenarios. We also apply the steering strategy to the large reasoning model, confirming its effectiveness in precise reasoning control.

Key facts. Cited by 24 papers in the map (data through 14 Oct 2026). Builds on 60 of its 80 references that are in the map. First posted on arXiv 23 May 2025.

Builds on

60 of 80 references.

References · 60

  1. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  2. 2020Measuring Massive Multitask Language Understandingcited by 7,963
  3. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,720
  4. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  5. 2023A Survey of Large Language Modelscited by 2,576
  6. 2023Sparse Autoencoders Find Highly Interpretable Features in Language Modelscited by 1,420
  7. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,255
  8. 2021Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivitycited by 1,135
  9. 2023Steering Llama 2 via Contrastive Activation Additioncited by 1,085
  10. 2024Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMscited by 570
  11. 2024Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Modelscited by 435
  12. 2024Mechanistic Interpretability for AI Safety -- A Reviewcited by 422
38 more
  1. 2021Attention is Not All You Need: Pure Attention Loses Rank Doubly Exponentially with Depthcited by 419
  2. 2024Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2cited by 394
  3. 2023Editing Large Language Models: Problems, Methods, and Opportunitiescited by 372
  4. 2021Do Prompt-Based Models Really Understand the Meaning of their Prompts?cited by 326
  5. 2019Intrinsic dimension of data representations in deep neural networkscited by 282
  6. 2020AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Promptscited by 249
  7. 2023PromptRobust: Towards Evaluating the Robustness of Large Language Models on Adversarial Promptscited by 226
  8. 2025AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoderscited by 218
  9. 2023Knowledge Editing for Large Language Models: A Surveycited by 194
  10. 2025Open Problems in Mechanistic Interpretabilitycited by 179
  11. 2025Thoughts Are All Over the Place: On the Underthinking of o1-Like LLMscited by 163
  12. 2024A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoderscited by 160
  13. 2025The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Taskscited by 152
  14. 2024Analyzing the Generalization and Reliability of Steering Vectorscited by 143
  15. 2024Llama Scope: Extracting Millions of Features from Llama-3.1-8B with Sparse Autoencoderscited by 143
  16. 2025SAEBench: A Comprehensive Benchmark for Sparse Autoencoders in Language Model Interpretabilitycited by 124
  17. 2024Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimizationcited by 118
  18. 2024Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Modelscited by 115
  19. 2025Are Sparse Autoencoders Useful? A Case Study in Sparse Probingcited by 115
  20. 2024Improving Steering Vectors by Targeting Sparse Autoencoder Featurescited by 92
  21. 2024What Did I Do Wrong? Quantifying LLMs' Sensitivity and Consistency to Prompt Engineeringcited by 90
  22. 2025Sparse Autoencoders Do Not Find Canonical Units of Analysiscited by 88
  23. 2024Detoxifying Large Language Models via Knowledge Editingcited by 87
  24. 2025AnyEdit: Edit Any Knowledge Encoded in Language Modelscited by 76
  25. 2023Word Embeddings Are Steers for Language Modelscited by 70
  26. 2025SEAL: Steerable Reasoning Calibration of Large Language Models for Freecited by 68
  27. 2024Steering Knowledge Selection Behaviours in LLMs via SAE-Based Representation Engineeringcited by 63
  28. 2025Trading Inference-Time Compute for Adversarial Robustnesscited by 61
  29. 2025Steering Large Language Model Activations in Sparse Spacescited by 58
  30. 2024Style Vectors for Steering Generative Large Language Modelcited by 56
  31. 2024Measuring and Controlling Instruction (In)Stability in Language Model Dialogscited by 53
  32. 2024The Geometry of Concepts: Sparse Autoencoder Feature Structurecited by 52
  33. 2024Knowledge Mechanisms in Large Language Models: A Survey and Perspectivecited by 51
  34. 2024Steering Without Side Effects: Improving Post-Deployment Control of Language Modelscited by 46
  35. 2024Why Larger Language Models Do In-context Learning Differently?cited by 45
  36. 2024Talking Heads: Understanding Inter-layer Communication in Transformer Language Modelscited by 41
  37. 2024Extending Activation Steering to Broad Skills and Multiple Behaviourscited by 40
  38. 2024Multi-property Steering of Large Language Models with Dynamic Activation Compositioncited by 39

50 most cited of 60 All 60 (JSON)

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 6; 2026: 18.

24 papers.

Sort:

By area:

  1. 2026Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Modelscited by 27
  2. 2025Steering Language Models with Weight Arithmeticcited by 20
  3. 2025SAGE: An Agentic Explainer Framework for Interpreting SAE Features in Language Modelscited by 10
  4. 2025AbsTopK: Rethinking Sparse Autoencoders For Bidirectional Featurescited by 9
  5. 2026Why Steering Works: Toward a Unified View of Language Model Parameter Dynamicscited by 9
  6. 2026From Data to Behavior: Predicting Unintended Model Behaviors Before Trainingcited by 8
  7. 2026Fine-Grained Activation Steering: Steering Less, Achieving Morecited by 7
  8. 2026Sparse Autoencoders are Capable LLM Jailbreak Mitigatorscited by 6
  9. 2026Steering at the Source: Style Modulation Heads for Robust Persona Controlcited by 5
  10. 2026ILRR: Inference-Time Steering Method for Masked Diffusion Language Modelscited by 4
  11. 2025Toward Preference-aligned Large Language Models via Residual-based Model Steeringcited by 3
  12. 2026How Controllable Are Large Language Models? A Unified Evaluation across Behavioral Granularitiescited by 3
12 more
  1. 2026Preference Heads in Large Language Models: A Mechanistic Framework for Interpretable Personalizationcited by 2
  2. 2025Towards Low-Resource Alignment to Diverse Perspectives with Sparse Feedbackcited by 1
  3. 2026C-DeltaTheta: Circuit-Restricted Weight Arithmetic for Selective Refusalcited by 1
  4. 2026Palette: A Modular, Controllable, and Efficient Framework for On-demand Authorized Safety Alignment Relaxation in LLMscited by 1
  5. 2026REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Featurescited by 1
  6. 2026Enhancing SAE-based Steering via Neighbor Integrated Feature Selectioncited by 1
  7. 2025When Models Refuse: Political Steerability and Feature Richness as Measures of Ideological Depthcited by 0
  8. 2026Activation Steering for Synthetic Data Generation: The Role of Diversity in Downstream Safety Detectioncited by 0
  9. 2026Who Bridges Safety? Identifying and Targeting Cross-Lingual Shared Safety Pathwayscited by 0
  10. 2026Key Path Identification for Resolving Knowledge Conflicts via SAE-based Steeringcited by 0
  11. 2026SafeTune: A Unified Faithful Library for Auditing and Repairing Safety Drift in Fine-Tuned LLMscited by 0
  12. 2026Query-Adaptive Neuron Steering for LLMscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Controllable LLM Reasoning via Sparse Autoencoder-Based Steering
  2. 2026Reliable Control-Point Selection for Steering Reasoning in Large Language Models
  3. 2025The Rogue Scalpel: Activation Steering Compromises LLM Safety
  4. 2025Interpretable LLM Guardrails via Sparse Representation Steering
  5. 2024Steering Language Model Refusal with Sparse Autoencoders
  6. 2025Feature-Guided SAE Steering for Refusal-Rate Control using Contrasting Prompts
  7. 2026A Coin Flip Per Token: Bernoulli Sparse Steering of Large Language Models
  8. 2026Steered Generation via Gradient-Based Optimization on Sparse Query Features