20 results for “activation steering”
CS papers onlyHybrid search: Keyword + semantic, ranked by combined score.ⓘ
Want pure semantic search? Try claim verification →
This paper systematically audits the safety implications of activation steering vectors, finding that these vectors significantly influence the success rate of jailbreak attacks by overlapping with la…
This paper proposes Policy Gradient Steering (PGS), a new method for dynamically changing a large language model's behavior at inference time using reinforcement learning.
Ioannis Prokopiou, Pantelis Vikatos, Maximos Kaliakatsos-Papakostas, Theodoros Giannakopoulos +1 more
This paper proposes a framework for deterministic attribute modulation in transformer-based architectures.
Luoyu Chen, Weiqi Wang, Zhiyi Tian, Chenhan Zhang +4 more
The paper proposes an unsupervised bi-level adversarial training framework to enhance LLM safety steering, achieving strong zero-shot defense against unseen and evolving jailbreak prompts.
Linfeng Liu, Tiffany Zhan, Louie Hong Yao, Saptarshi Ghosh +1 more
The paper demonstrates that the internal signals governing figurative language generation are reusable across multiple languages, showing that a steering direction learned in one language can effectiv…
Jona te Lintelo, Lichao Wu, Marina Krček, Sengim Karayalçin +1 more
MASCing is a novel framework that enables flexible, non-retraining reconfiguration of Mixture-of-Experts (MoE) models for specific safety objectives by applying activation steering masks to control ex…
Evgenii Kortukov, Piotr Komorowski, Florian Klein, Paula Engl +4 more
This paper introduces a new method for controlling the behavior of large reasoning models with minimal output quality degradation.
Ioannis Prokopiou, Pantelis Vikatos, Maximos Kaliakatsos-Papakostas, Theodoros Giannakopoulos +1 more
The paper proposes an inference-time activation steering framework, utilizing orthogonalization, to achieve fine-grained, deterministic control over discrete musical attributes like Pitch and Duration…
Yu Yan, Jiahao Chen, Siqi Lu, Yongjuan Wang +3 more
This paper investigates security risks in Steering plugins for large language models and proposes a malicious plugin and a runtime defense framework to counter such threats.
The paper proposes SafeDIG, a robust safety steering framework that adapts Diffusion Transformers for text-to-image generation by treating safety control as position-aware sparse feature transfer, ens…
This paper analyzes the internal mechanism of full-duplex spoken language models and proposes a method to improve their ability to handle interruptions.
This paper introduces instruction-based vector steering to redistribute attention in Large Audio-Language Models (LALMs) and demonstrates its effectiveness in recovering sound event locations.
This paper demonstrates that Sparse Autoencoders (SAEs) can effectively steer Large Language Models (LLMs) on the AxBench benchmark, achieving performance comparable to LoRA baselines when combined wi…
The paper proposes CTRL-STEER, a closed-loop framework that adaptively adjusts intervention strength to stabilize concept regulation and improve task success in Vision-Language-Action models without r…
The paper introduces an adaptive probe-based steering method that significantly improves the robustness and effectiveness of LLM jailbreaking without requiring extra prompts or manual tuning.
Andy Tang, William Chen, Andrew Wagenmaker, Chelsea Finn +1 more
This paper proposes a method called Flow Reversal Steering (FRS) to improve generalist policies in robots by inferring and invoking appropriate actions from their behavioral prior.
Jonathan DeCastro, Thomas Balch, Emily Sumner, Deepak Gopinath +8 more
The paper presents a model-based reinforcement learning framework for assistive agents in shared-control settings that addresses the challenges of proactive assistance and human legibility.
Zehan Zhang, Neng Zhang, Yaoyi Li, Jia Cai +1 more
The paper proposes a diffusion-based planning framework called Diffusion Forcing Planner (DFP) to stabilize motion plans in autonomous driving.
The paper shows that safety failures in low-resource languages are due to a failure in the model's safety decision calibration, not a lack of underlying knowledge, and proposes a recalibration method…