COSMIC: Generalized Refusal Direction Identification in LLM Activations (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods depend on predefined refusal templates detectable in output tokens or manual review. |
| Approach: | They propose a framework that optimally identifies steering directions and target layers using cosine similarity, entirely independent of output text. |
| Outcome: | The proposed framework achieves comparable steering effectiveness without any prior knowledge or assumptions of a model’s refusal behavior such as the use of certain refusal tokens. |
Similar Papers
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. |
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. |
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. |
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 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. |
Mitigating Over-Refusal in Aligned Large Language Models via Inference-Time Activation Energy (2026.acl-long)
Copied to clipboard
Eric Hanchen Jiang, Weixuan Ou, Run Liu, Shengyuan Pang, Guancheng Wan, Ranjie Duan, Wei Dong, Kai-Wei Chang, XiaoFeng Wang, Ying Nian Wu, Xinfeng Li
| Challenge: | Existing safety alignment techniques prioritize mitigating harmful responses at the expense of overcautious behavior, leading models to incorrectly refuse benign requests. |
| Approach: | They propose a fine-tuning free framework to improve safety and reduce false refusals by dynamic, inference-time intervention. |
| Outcome: | The proposed framework raises compliance on the ORB-H benchmark from 57.3% to 82.6% while maintaining the baseline safety performance. |
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. |
The Unintended Trade-off of AI Alignment: Balancing Hallucination Mitigation and Safety in LLMs (2026.findings-eacl)
Copied to clipboard
| Challenge: | Hallucination in large language models has been studied, but a side effect remains unrecognized . a new study examines the trade-off between truthfulness and safety alignment . |
| Approach: | They propose a method that disentangles hallucination from hallucinian features using sparse autoencoders. |
| Outcome: | The proposed method preserves refusal behavior and task utility while maintaining safety alignment. |
SafeConstellations: Mitigating Over-Refusals in LLMs Through Task-Aware Representation Steering (2026.acl-long)
Copied to clipboard
| Challenge: | Current safety alignment methods fail to identify intended benign task before refusing to respond. |
| Approach: | They propose a method that uses inference-time trajectory-shifting to guide model behavior . they show that LLMs persist in refusing inputs containing harmful content . |
| Outcome: | The proposed approach reduces over-refusals with minimal impact on utility. |
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. |