Large Language Models (LLMs) are commonly finetuned for a variety of use cases and domains. A common approach is to leverage Low-Rank Adaptation (LoRA) -- known to provide strong performance at low resource costs. In this study, we demonstrate that LoRA actually opens the door to short-cut vulnerabilities -- and the more resource efficient is the LoRA setup, the more vulnerable will be the finetuned model to aggressive attacks. To measure that vulnerability, we introduce Seamless Spurious Token Injection (SSTI), where we find that LoRA exclusively focuses on even just a single token that is spuriously correlated with downstream labels. In short, injection of that spurious token during finetuning ensure that the model's prediction at test-time can be manipulated on-demand. We conducted experiments across model families and datasets to evaluate the impact of SSTI during LoRA finetuning while providing possible mitigations. Our experiments conclude that none of the existing checkers and preprocessors can sanitize a dataset raising new concerns for data quality and AI safety.

In recent years, the importance of developing targeted strategies to pursue opportunities in the Asia Pacific – that is, 'Asia Strategies' – has come to the forefront of policy discourse.[1] As noted in global strategy advisor Parag Khanna's bestseller, The Future is Asian, Asia is increasingly becoming more prosperous and confident. For Canada, sustained engagement with this region is now critical to the country's long-term economic prosperity. Propelled by its young population and increase in productivity, Asia is predicted to make up 65 per cent of the world's global middle class by 2030 and it is expected to account for 53 per cent of the world's population and 52 per cent of global GDP by 2050. The need for trade diversification and deeper engagement with Asia is clear, but Canada often falls short of the Asia-specific skills, resources, and, most importantly, strategiesthat are required to take advantage of these opportunities.