RealControlBench: A Benchmark for Interpretability-Informed Fine-Grained Control of LLMs and Agents

Xinying Cai, Bin Hu, Mingyu Jin, Qiming Zhang, Yang Ouyang, Bangwei Guo, Minghao Guo, Yutong Yin, Lang Gao, Feijiang Han, Sheldon Yu, Kaiqu Liang, Qingcheng Zeng, Jingcheng Niu, Fan Yang, Wei Cheng, Zhen Xiang, Julian McAuley and Zhaoran Wang.
arXiv preprint

TL;DR

Existing steering methods remain unreliable once control objectives look like real deployment. RealControlBench moves steering evaluation beyond isolated, single-attribute generations to execution-grounded, multi-turn agent interactions across five dimensions: instruction-following control, agentic tool-use decision control, reasoning-performance steering, reasoning-length control, and trajectory-conditioned safety awareness. It measures both target controllability and capability preservation, and finds that current methods often substantially degrade task performance.

How to Cite

@article{cai2026realcontrolbench,
  title   = {{RealControlBench}: A Benchmark for Interpretability-Informed Fine-Grained Control of {LLM}s and Agents},
  author  = {Cai, Xinying and Hu, Bin and Jin, Mingyu and Zhang, Qiming and Ouyang, Yang and Guo, Bangwei and Guo, Minghao and Yin, Yutong and Gao, Lang and Han, Feijiang and Yu, Sheldon and Liang, Kaiqu and Zeng, Qingcheng and Niu, Jingcheng and Yang, Fan and Cheng, Wei and Xiang, Zhen and McAuley, Julian and Wang, Zhaoran},
  journal = {arXiv preprint arXiv:2610.27295},
  year    = {2026},
}