Selective Steering: Norm-Preserving Control Through Discriminative Layer Selection (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for inference-time steering are limited by their limitations . Angular Steering violates norm preservation, causing distribution shift and generation collapse . |
| Approach: | They propose a method that uses a norm-preserving rotation formulation to maintain activation distribution integrity and discriminative layer selection to apply steering only where features exhibit opposite-signed class alignment. |
| Outcome: | Experiments show that Selective Steering achieves higher attack success rates than prior methods while maintaining zero perplexity violations and approximately 100% capability retention on standard benchmarks. |
Similar Papers
Beyond Linear Steering: Unified Multi-Attribute Control for Language Models (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Empirical results show that K-Steering outperforms strong baselines in accurately steering multiple behaviors. |
| Approach: | They propose a method that trains a single classifier on hidden activations and computes intervention directions via gradients at inference time. |
| Outcome: | The proposed method outperforms strong baselines in steering multiple behaviors. |
Steering Safely or Off a Cliff? Rethinking Specificity and Robustness in Inference-Time Interventions (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing studies have shown that model steering can preserve fluency and unrelated abilities, but it fails to preserve robustness specificity. |
| Approach: | They propose a framework that distinguishes three dimensions of specificity: general, control, and robustness. |
| Outcome: | The proposed framework distinguishes three dimensions of specificity: general (preserving fluency and unrelated abilities), control (preserving related control properties), and robustness (preserving control properties under distribution shifts). |
Shifting Perspectives: Steering Vectors for Robust Bias Mitigation in LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Despite efforts to mitigate social bias in large language models, representational harms such as stereotyping continue to exist in both open and closed-source models. |
| Approach: | They propose a method to modify model activations in forward passes by applying steering vectors to a BBQ dataset and comparing their results to bias mitigation methods. |
| Outcome: | The proposed method outperforms 3 other bias mitigation methods on the BBQ dataset and shows the lowest impact on MMLU scores. |
Representation Bending for Large Language Model Safety (2025.acl-long)
Copied to clipboard
Ashkan Yousefpour, Taeheon Kim, Ryan Sungmo Kwon, Seungbeen Lee, Wonje Jeung, Seungju Han, Alvin Wan, Harrison Ngan, Youngjae Yu, Jonghyun Choi
| Challenge: | Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and fail to generalize across unseen attacks. |
| Approach: | They propose a new approach that disrupts representations underlying harmful behaviors in Large Language Models by using loss-based fine-tuning. |
| Outcome: | The proposed approach outperforms existing methods such as Circuit Breaker, RMU, and NPO with 95% reduction in attack success rates across diverse jailbreak benchmarks. |
AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender (2025.emnlp-main)
Copied to clipboard
Weixiang Zhao, Jiahe Guo, Yulin Hu, Yang Deng, An Zhang, Xingyu Sui, Xinyang Han, Yanyan Zhao, Bing Qin, Tat-Seng Chua, Ting Liu
| Challenge: | Activation steering offers training-free defense but relies on fixed steering coefficients, resulting in suboptimal protection and increased false rejections of benign inputs. |
| Approach: | They propose an adaptive activation steering method that dynamically adjusts model behavior based on input characteristics. |
| Outcome: | The proposed method outperforms baseline methods across multiple jailbreak attacks with minimal impact on utility. |
CogSteer: Cognition-Inspired Selective Layer Intervention for Efficiently Steering Large Language Models (2025.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) achieve excellent performance through pretraining on extensive data. |
| Approach: | They propose an efficient selective layer intervention based on parameter-efficient fine-tuning methods to select the optimal steering layer to modulate LLM semantics. |
| Outcome: | The proposed approach is based on a model-agnostic framework and is safe to deploy. |
Activation-Space Personality Steering: Hybrid Layer Selection for Stable Trait Control in LLMs (2026.eacl-long)
Copied to clipboard
| Challenge: | Personality-aware LLMs exhibit implicit personalities in their generation, but reliably controlling or aligning these traits to meet specific needs remains an open challenge. |
| Approach: | They propose a pipeline that extracts hidden state activations from transformer layers using the Big Five Personality Traits framework. |
| Outcome: | The proposed model extracts hidden state activations from transformer layers using the Big Five personality traits (Openness, Conscientiousness, Extraversion, Agreeableness and Neuroticism) |
Detecting What Queries Seek: Steering LLM Safety with FFN Output Activation Monitoring (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for detecting malicious queries rely on residual stream activations, resulting in limited discriminative power and unreliable interventions. |
| Approach: | They propose to use feed-forward networks (FFNs) to generate more discriminative signals for intervention, since these activations more explicitly reflect the intent of a query. |
| Outcome: | Experiments show that the proposed approach achieves state-of-the-art defense performance against various jailbreak attacks while maintaining the model's original performance on benign tasks. |
Activation Decomposition and Steering for LLM Backdoor Remediation (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to defending against LLM backdoors rely on auxiliary models or safety-related datasets. |
| Approach: | They propose a method which contrasts benign and poisoned settings to decompose feature vectors for steering without auxiliary models or datasets. |
| Outcome: | The proposed method achieves better defense qualities than existing steering strategies. |
Householder Pseudo-Rotation: A Novel Approach to Activation Editing in LLMs with Direction-Magnitude Perspective (2024.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to edit LLMs' activations are limited by their magnitude and direction consistency. |
| Approach: | They propose a method that edits activations to alter their magnitudes and directions to preserve activation norms. |
| Outcome: | The proposed method preserves activation norm and improves safety benchmarks. |