ACL, The 64th Annual Meeting of the Association for Computational Linguistics
We are thrilled to congratulate the team from the University of Liverpool for winning the ๐ฝ๐ฟ๐ฒ๐๐๐ถ๐ด๐ถ๐ผ๐๐ ๐ข๐๐๐๐๐ฎ๐ป๐ฑ๐ถ๐ป๐ด ๐ฃ๐ฎ๐ฝ๐ฒ๐ฟ ๐๐๐ฎ๐ฟ๐ฑ at the 64th Annual Meeting of the Association for Computational Linguistics (ACL 2026), held in San Diego, United States, from July 2 to 7, 2026.
Their award-winning paper, “๐๐๐ถ๐ป๐ด ๐๐ถ๐๐ต ๐ง๐ฟ๐๐๐ต๐: ๐ข๐ฝ๐ฒ๐ป-๐๐ต๐ฎ๐ป๐ป๐ฒ๐น ๐ ๐๐น๐๐ถ-๐๐ด๐ฒ๐ป๐ ๐๐ผ๐น๐น๐๐๐ถ๐ผ๐ป ๐ณ๐ผ๐ฟ ๐๐ฒ๐น๐ถ๐ฒ๐ณ ๐ ๐ฎ๐ป๐ถ๐ฝ๐๐น๐ฎ๐๐ถ๐ผ๐ป ๐๐ถ๐ฎ ๐๐ฒ๐ป๐ฒ๐ฟ๐ฎ๐๐ถ๐๐ฒ ๐ ๐ผ๐ป๐๐ฎ๐ด๐ฒ,” was authored by Jinwei Hu , Xinmiao Huang, Yi Dong, and Xiaowei Huang, from the School of Computer Science and Informatics at the University of Liverpool, in collaboration with Youcheng Sun from the MBZUAI (Mohamed bin Zayed University of Artificial Intelligence).
๐ช๐ต๐ฎ๐ ๐ถ๐ ๐๐ต๐ถ๐ ๐ฟ๐ฒ๐๐ฒ๐ฎ๐ฟ๐ฐ๐ต ๐ฎ๐ฏ๐ผ๐๐?
As Large Language Models LLMs increasingly operate as the core of autonomous agents and synthesize information from multiple sources, their tendency to construct coherent narratives can become a new attack surface.
Inspired by cinematic montage, where the arrangement of separate shots can create new and potentially misleading meanings, the authors introduce cognitive collusion: multiple agents can manipulate a victim agent using only truthful fragmented information โ without fabricated information, hidden channels, or backdoors. By strategically selecting and arranging these fragments, attackers can lead the model to form and believe a convincing but false conclusion.
They further introduce Generative Montage, a multi-agent framework for constructing and distributing such deceptive narratives based on real-world rumor events. Experiments across 14 LLM families show that this attack is highly effective, and that stronger reasoning capabilities may sometimes make models even more vulnerable and overconfident.
These findings highlight the need for AI safeguards that examine not only whether individual pieces of information are true, but also how they are combined and interpreted.
You can find the whole presentation made by Jinwei Hu and the poster: https://robustifai.eu/dissemination/
Congratulations to the whole team on this well-deserved recognition at ACL 2026!
ย




