analiza aiqbench
This research introduces CAM-Steer, a novel activation steering framework designed to simultaneously mitigate multiple harm categories within large language models during inference. By dynamically coordinating safety directions and intervention strengths based on estimated risks across different threat types, the method effectively handles complex prompts where single-category safeguards often fail. The approach demonstrates superior defense success rates with negligible overhead, offering a practical pathway toward more robust and comprehensive LLM safety protocols.
How to Tame a Multi-Headed Hydra? Adaptive Multi-Category Safety Steering for Large Language Models
arXiv:2609.34514v1 Announce Type: cross Abstract: As large language models (LLMs) become increasingly widespread, preventing unsafe responses to harmful prompts is essential for their safe deployment. Activation steering offers an approach to improving LLM safety by modifying internal activations during inference without updating model parameters. However, a single prompt can involve multiple harm categories, and steering toward safety in one category may leave harmful content from another unaddressed. Despite advances in adaptive steering, existing methods do not explicitly coordinate steering direction and strength when multiple harm categories co-occur within a single prompt. To address this problem, we propose CAM-Steer, a Category-Adaptive Multi-category Safety Steering framework. Specifically, it estimates the risk associated with each harm category by comparing the current hidden state with safe and unsafe prototypes. The estimated risks are then used to combine the safety directions for different harm categories into a single steering direction and to determine the strength of the intervention. Finally, it rotates the hidden state along the composed steering direction, with the rotation angle determined by the estimated risks, while preserving the hidden-state norm. Experiments across three LLM backbones and seven harm categories show that CAM-Steer outperforms the evaluated baselines in average defense success rate, including when categories co-occur. Further analyses support its component designs and informative risk scores, with negligible inference overhead.
Title:How to Tame a Multi-Headed Hydra? Adaptive Multi-Category Safety Steering for Large Language Models
Abstract:As large language models (LLMs) become increasingly widespread, preventing unsafe responses to harmful prompts is essential for their safe deployment. Activation steering offers an approach to improving LLM safety by modifying internal activations during inference without updating model parameters. However, a single prompt can involve multiple harm categories, and steering toward safety in one category may leave harmful content from another unaddressed. Despite advances in adaptive steering, existing methods do not explicitly coordinate steering direction and strength when multiple harm categories co-occur within a single prompt. To address this problem, we propose CAM-Steer, a Category-Adaptive Multi-category Safety Steering framework. Specifically, it estimates the risk associated with each harm category by comparing the current hidden state with safe and unsafe prototypes. The estimated risks are then used to combine the safety directions for different harm categories into a single steering direction and to determine the strength of the intervention. Finally, it rotates the hidden state along the composed steering direction, with the rotation angle determined by the estimated risks, while preserving the hidden-state norm. Experiments across three LLM backbones and seven harm categories show that CAM-Steer outperforms the evaluated baselines in average defense success rate, including when categories co-occur. Further analyses support its component designs and informative risk scores, with negligible inference overhead.
Cryptography and Security (cs.CR); Computation and Language (cs.CL); Machine Learning (cs.LG)
arXiv:2609.34514 [cs.CR]
(or arXiv:2609.34514v1 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2609.34514 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 328 cuvinte
Subiecte conexe
[AINews] Reality Checks on AI News (Yegge shuts down Gas Town, Databricks’ +60% Astra cost)
Latent Space · 17 sept. 2026
[AINews] High Taste Testers Ep. 1
Latent Space · 17 oct. 2026
Introducing CARE-X: Towards Clinically Useful Radiology VLMs with Auxiliary Supervision, Reward-Aligned Learning, and Tool-Augmented Measurement
Microsoft Research · 11 aug. 2026
Angles of attack for continual learning safety
LessWrong · 16 iun. 2026