Multimodal large language models (MLLMs) have shown impressive capabilities in vision-language tasks such as reasoning segmentation, where models generate segmentation masks based on textual queries. While prior work has primarily focused on perturbing image inputs, semantically equivalent textual paraphrases—crucial in real-world applications where users express the same intent in varied ways—remain underexplored. To address this gap, we introduce a novel adversarial paraphrasing task: generating grammatically correct paraphrases that preserve the original query meaning while degrading segmentation performance. To evaluate the quality of adversarial paraphrases, we develop a comprehensive automatic evaluation protocol validated with human studies. Furthermore, we introduce SPARTA—a black-box, sentence-level optimization method that operates in the low-dimensional semantic latent space of a text autoencoder, guided by reinforcement learning. SPARTA achieves significantly higher success rates, outperforming prior methods by up to 2x on both the ReasonSeg and LLMSeg-40k datasets. We use SPARTA and competitive baselines to assess the robustness of advanced reasoning segmentation models. We reveal that they remain vulnerable to adversarial paraphrasing—even under strict semantic and grammatical constraints. All code and data will be released publicly upon acceptance.

This paper presents a model for solving the Multiple Choice Question Answering (MCQA) problem, focusing on the impact of subgraph extraction from a Knowledge Graph on model performance. The proposed method combines textual and graph information by adding linearized subgraphs directly into the main question prompt with separate tokens, enhancing the performance of models working with each modality separately. The study also includes an examination of Large Language Model (LLM) backbones and the benefits of linearized subgraphs and sequence length, with efficient training achieved through fine-tuning with LoRA. The top benchmark, using subgraphs and MPNet, achieved an F1 score of 0.3887. The main limitation of the experiments is the reliance on pre-generated subgraphs/triplets from the graph, and the lack of exploration of in-context learning and prompting strategies with decoder-based architectures.