Authors: W Liu, X Mou, H Yan, Z, Wei, Y He
Year: 2026
Published in: arXiv preprint arXiv:2606.04075, 2026•arxiv.org
Institution: King’s College London, Fudan University, Shanghai Innovation Institute, The Alan Turing Institute
Research Area: Human-Computer Interaction
Discipline: Machine Learning, Artificial Intelligence
The paper finds that large language models can exploit gaps in societal rules, leading to regulatory loophole discovery, necessitating a new post-training approach for safely integrating LLMs into society.
Methods: The study introduced the SocioHack sandbox, consisting of 72 societal environments, to investigate reward hacking and loophole discovery by LLMs.
Key Findings: The study measured the emergence of reward hacking in societal environments and the ability of models to find and exploit loopholes in social rules.
Sample Size: 72
Authors: Z Qiu, W Liu, H Feng, Z Liu, T Xiao
Year: 2024
Published in: ArXiv
Institution: Massachusetts Institute of Technology, Max Planck Institute, University of Cambridge
Research Area: Computational cognition, LLM evaluation, Program synthesis, Multimodal reasoning
Discipline: Artificial Intelligence
Introduces SGP-Bench, a benchmark testing whether LLMs can answer semantic and spatial questions about images purely from graphics programs (SVG/CAD), effectively probing “visual imagination without vision.” The authors show current LLMs struggle - sometimes near chance - even when images are trivial for humans, but demonstrate that Symbolic Instruction Tuning (SIT) can meaningfully improve thi...
