Harmful meme detection requires a stable understanding of how the image and the text work together. Since the text is already given as external input, a natural test is whether predictions remain stable when the corresponding visible text is removed from the image. Motivated by test-time multi-view inference, we build a text-erased reference view by keeping the external text unchanged and removing the visible text. If the model uses the external text in a stable way, this change should not greatly affect its harmfulness prediction. However, we observe clear score shifts and prediction flips, showing cross-view instability under fixed external text. We further find that the cross-view change after neutral-probe filtering is related to error risk, so it can serve as a test-time diagnostic signal. Based on this finding, we propose TIDC, a test-time framework that uses the filtered signal to calibrate the original-view score. We also provide theoretical analysis to support the diagnostic signal and calibration design. Experiments on HarM, FHM, and MAMI show that TIDC improves zero-shot performance across frozen multimodal backbones, without parameter updates or reasoning chains.

Large Multimodal Models (LMMs) exhibit impressive cross-modal understanding and reasoning abilities, often assessed through multiple-choice questions (MCQs) that include an image, a question, and several options. However, many benchmarks used for such evaluations suffer from systematic biases. Remarkably, Large Language Models (LLMs) without any visual perception capabilities achieve non-trivial performance, undermining the credibility of these evaluations. To address this issue while maintaining the efficiency of MCQ evaluations, we propose MMEVALPRO, a benchmark designed to avoid Type-I errors through a trilogy evaluation pipeline and more rigorous metrics. For each original question from existing benchmarks, human annotators augment it by creating one perception question and one knowledge anchor question through a meticulous annotation process. MMEVALPRO comprises 2,138 question triplets, totaling 6,414 distinct questions. Two-thirds of these questions are manually labeled by human experts, while the rest are sourced from existing benchmarks (MMMU, ScienceQA, and MathVista). Compared with the existing benchmarks, our experiments with the latest LLMs and LMMs demonstrate that MMEVALPRO is more challenging (the best LMM lags behind human performance by 31.73%, compared to an average gap of 8.03% in previous benchmarks) and more trustworthy (the best LLM trails the best LMM by 23.09%, whereas the gap for previous benchmarks is just 14.64%). Our in-depth analysis explains the reason for the large performance gap and justifies the trustworthiness of evaluation, underscoring its significant potential for advancing future research.

Molecular representation learning plays an essential role in cheminformatics. Recently, language model-based approaches have gained popularity as an alternative to traditional expert-designed features to encode molecules. However, these approaches only utilize a single molecular language for representation learning. Motivated by the fact that a given molecule can be described using different languages such as Simplified Molecular Line Entry System (SMILES), The International Union of Pure and Applied Chemistry (IUPAC), and The IUPAC International Chemical Identifier (InChI), we propose a multilingual molecular embedding generation approach called MM-Deacon (multilingual molecular domain embedding analysis via contrastive learning). MM-Deacon is pre-trained using SMILES and IUPAC as two different languages on large-scale molecules. We evaluated the robustness of our method on seven molecular property prediction tasks from MoleculeNet benchmark, zero-shot cross-lingual retrieval, and a drug-drug interaction prediction task.