The paper introduces Minimally Invasive Steering Vector Optimization (MISVO), a method that adjusts a frozen language model’s final hidden states by adding vectors to steer outputs toward a test‑time reward while penalizing changes using the local KL geometry of the token distribution. MISVO derives an analytic Fisher term and a suffix score‑function term, showing that the suffix term is second‑order and that Fisher surrogates match the full KL gradient to first order. Experiments on preference and code‑generation tasks with 1B–14B parameter models demonstrate that MISVO achieves the highest mean reward in most settings while maintaining diversity and coherence comparable to Best‑of‑N.
By Taha Entesari, Jingyu Zhang, Daniel Khashabi, Mahyar Fazlyab
arXiv:2605. 05983v2 Announce Type: replace Abstract: Recently, steering vectors (SVs) have emerged as an effective and lightweight approach to steer behaviors of large language models (LLMs), among which fine-tuned SVs are more effective than optimization-free ones.
By Yuntai Bao, Qinfeng Li, Xinyan Yu, Ge Su, Wenqi Zhang, Liu Yan, Haiqin Weng, Jianwei Yin, Xuhong Zhang
arXiv:2606. 11599v1 Announce Type: cross Abstract: Activation steering offers a lightweight approach to control language models' behavior at inference time, but whether it succeeds or fails heavily depends on the prompt, concept, model, and steering configuration.
By Chenrui Fan, Yize Cheng, Ming Li, Soheil Feizi, Tianyi Zhou
The paper challenges the notion that refusal in large language models is governed by a single direction in activation space. It demonstrates that different refusal and non‑compliance categories map to distinct geometric directions, yet steering along any of these directions yields similar refusal–over‑refusal trade‑offs, acting as a shared one‑dimensional control knob. Using sparse autoencoders, the authors reveal a structured internal representation of refusal, comprising a reusable core of shared latents and style‑ or domain‑specific latents, and show that linear interventions collapse this structure into uniform behavioral control.
By Faaiz Joad, Majd Hawasly, Sabri Boughorbel, Nadir Durrani, Husrev Taha Sencar
arXiv:2607. 25907v1 Announce Type: cross Abstract: Activation steering controls model behavior by editing internal activations at inference time.
By Deepanshu Mody, Samarth Agarwal, Utkarsh Mittal, Dipesh Mahato
arXiv:2605.16339v2 Announce Type: replace
Abstract: Preference learning in large language models relies on reward models as proxies for human judgment. However, these models frequently exhibit prefer...
By Shunchang Liu, Xin Chen, Belen Martin Urcelay, Francesco Croce
arXiv:2602. 02712v2 Announce Type: replace Abstract: A popular approach to post-training control of large language models (LLMs) is the steering of intermediate latent representations.
By Magamed Taimeskhanov, Samuel Vaiter, Damien Garreau
arXiv:2607. 25270v1 Announce Type: cross Abstract: Activation steering controls language models by adding vectors or features to hidden states at inference time, but the upstream source of these steering signals is often treated as a secondary detail.
By Jiaran Ye, Lingxu Ran, Zijun Yao, Chenpeng Wang, Yong Jiang, Lei Hou, Juanzi Li, Liangming Pan
arXiv:2606. 06673v1 Announce Type: new Abstract: Sparse rewards and heterogeneous task sequences remain persistent challenges in Reinforcement Learning (RL), often resulting in slow convergence, weak generalization, and inefficient exploration.
By Ujjwal Bhatta, Utsabi Dangol, Sumaly Bajracharya, Rodrigue Rizk, KC Santosh
Activation steering offers a lightweight approach to control language models' behavior at inference time, but whether it succeeds or fails heavily depends on the prompt, concept, model, and steering configuration. Finding the regime and boundaries of successful steering typically requires expensive grid searches and post-hoc evaluation of full autoregressive rollouts.
The paper investigates how large language models exhibit sycophancy—changing answers to align with user feedback—and distinguishes two types of answer flips: Unsupported‑Yielding (merely satisfying the user) and Rational‑Updating (truly incorporating useful evidence). Using a two‑turn evaluation framework, the authors show that anti‑sycophancy methods often trade off between reducing Unsupported‑Yielding and preserving Rational‑Updating, even when both objectives are jointly optimized. Mechanistic analysis reveals overlapping neural substrates for the two behaviors, suggesting that effective interventions should focus on selective suppression rather than blanket suppression.
By Huanhuan Ma, Henry Peng Zou, Chengze Li, Enze Ma, Yunyue Su, Philip S. Yu
arXiv:2606. 15092v1 Announce Type: new Abstract: Activation steering has emerged as a key methodology for controlling the behavior of large language models (LLMs).
By Minh-Hieu Pham, Bach Do, Laziz Abdullaev, Tan Minh Nguyen, Khoat Than