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. |
Similar Papers
Embracing Anisotropy: Turning Massive Activations into Interpretable Control Knobs for Large Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Prior work shows that Large Language Models exhibit highly anisotropic internal representations . prior work shows specialized dimensions capture domain-specific features . |
| Approach: | They propose a simple magnitude-based criterion to identify Domain-Critical Dimensions in a training-free manner. |
| Outcome: | The proposed method outperforms whole-dimension steering in domain adaptation and jailbreaking scenarios. |
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. |
Activation Scaling for Steering and Interpreting Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | a successful intervention should flip the correct with the wrong token, while remaining sparse. |
| Approach: | They propose to use activation scaling to flip the correct with the wrong token . they use gradient-based optimization to learn and evaluate a specific kind of efficient intervention . |
| Outcome: | The proposed method performs comparable with steering vectors but is much less minimal. |
Token-Aware Editing of Internal Activations for Large Language Model Alignment (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to optimize the behavior of large language models neglect misalignment discrepancies among tokens, resulting in deviant alignment direction and inflexible editing strength. |
| Approach: | They propose a token-aware editing approach to exploit the misalignment discrepancy among tokens to enhance activation probing and facilitate intervention. |
| Outcome: | Extensive experiments on three alignment capabilities demonstrate the efficacy of the proposed approach surpassing baseline by 25.8% on the primary metric of truthfulness with minimal cost. |
Style Vectors for Steering Generative Large Language Models (2024.findings-eacl)
Copied to clipboard
Kai Konen, Sophie Jentzsch, Diaoulé Diallo, Peer Schütt, Oliver Bensch, Roxanne El Baff, Dominik Opitz, Tobias Hecking
| Challenge: | Large language models (LLMs) can be trained on vast corpora and can generate text in a nuanced and parameterisable way. |
| Approach: | They propose to add style vectors to the activations of hidden layers during text generation to steer output towards specific styles. |
| Outcome: | The proposed approach differs from prompt engineering in that it can be nuanced and parameterisable. |
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. |
Activation Steering Decoding: Mitigating Hallucination in Large Vision-Language Models through Bidirectional Hidden State Intervention (2025.acl-long)
Copied to clipboard
| Challenge: | Large Vision Language Models (LVLMs) suffer from hallucination where generated textual descriptions fail to align accurately with visual semantics. |
| Approach: | They propose a training-free approach that mitigates hallucination through targeted intervention in the model’s intermediate activations by identifying directional patterns of hallucinism in the activation space using a small calibration set. |
| Outcome: | The proposed approach reduces hallucination across multiple benchmarks while maintaining performance on general visual understanding tasks. |
Internal Value Alignment in Large Language Models through Controlled Value Vector Activation (2025.acl-long)
Copied to clipboard
| Challenge: | Existing LLMs do not possess consistent values, but many have been developed to align them at the behavioral level, including supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). |
| Approach: | They propose a Controlled Value Vector Activation method that directly aligns the internal values of Large Language Models by interpreting how a value is encoded in their latent representations. |
| Outcome: | The proposed method achieves highest success rate across 10 basic values without hurting model performance and fluency, and ensures target values even with opposite and potentially malicious input prompts. |
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. |
Editing Large Language Models: Problems, Methods, and Opportunities (2023.emnlp-main)
Copied to clipboard
Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, Ningyu Zhang
| Challenge: | Recent advances in model editing for LLMs have created challenges and opportunities for the community. |
| Approach: | They propose to alter the behavior of LLMs efficiently within a specific domain without negatively impacting performance across other inputs. |
| Outcome: | The proposed method alters behavior of LLMs efficiently within a specific domain without negatively impacting performance across other inputs. |