Idiomatic expressions present a unique chal-lenge in NLP, as their meanings are often notdirectly inferable from their constituent words.Despite recent advancements in large languagemodels, idiomaticity remains a significant ob-stacle to robust semantic representation. Wepresent datasets and task results for MWE-2026 Shared Task 2: Advancing MultimodalIdiomaticity Representation 2 (AdMIRe 2),which challenges the community to assess andimprove models’ ability to interpret idiomaticexpressions in multimodal contexts across mul-tiple languages. Participants competed in animage ranking task in which, for each item,systems receive a context sentence containinga potentially idiomatic expression (PIE) andfive candidate images. Participating systemsare required to predict the sentence type (i.e.,idiomatic vs. literal) for the given context andrank the images by how well they depict the in-tended meaning in that context. Among the par-ticipating systems the most effective methodsinclude pipelines utilizing closed-source com-mercial models such as Gemini 2.5 and GPT-5, and employing chain-of-thought reasoningstrategies. Methods to mitigate language mod-els’ bias towards literal interpretations and en-sembles to smooth out variance were common.

Idiom corpora typically include both idiomatic and literal examples of potentially idiomatic expressions, but creating such corpora traditionally requires substantial expert effort and cost. In this article, we explore the use of large language models (LLMs) to generate synthetic idiom corpora as a more time- and cost-efficient alternative. We evaluate the effectiveness of synthetic data in training task-specific models and testing GPT-4 in few-shot prompting setting using synthetic data for idiomaticity detection. Our findings reveal that although models trained on synthetic data perform worse than those trained on human-generated data, synthetic data generation offers considerable advantages in terms of cost and time. Specifically, task-specific idiomaticity detection models trained on synthetic data outperform the general-purpose LLM that generated the data when evaluated in a zero-shot setting, achieving an average improvement of 11 percentage points across four languages. Moreover, synthetic data enhances the LLM’s performance, enabling it to match the task-specific models trained with synthetic data when few-shot prompting is applied.