Multimodal Large Language Models (MLLMs) have achieved remarkable progress in multimodal reasoning. However, they often excessively rely on textual information during the later stages of inference, neglecting the crucial integration of visual input. Current methods typically address this by explicitly injecting visual information to guide the reasoning process. In this work, through an analysis of MLLM attention patterns, we made an intriguing observation: with appropriate guidance, MLLMs can spontaneously re-focus their attention on visual inputs during the later stages of reasoning, even without explicit visual information injection. This spontaneous shift in focus suggests that MLLMs are intrinsically capable of performing visual fusion reasoning. Building on this insight, we introduce Look-Back, an implicit approach designed to guide MLLMs to ``look back" at visual information in a self-directed manner during reasoning. Look-Back empowers the model to autonomously determine when, where, and how to re-focus on visual inputs, eliminating the need for explicit model-structure constraints or additional input. We demonstrate that Look-Back significantly enhances the model's reasoning and perception capabilities, as evidenced by extensive empirical evaluations on multiple multimodal benchmarks.

The use of machine-learning techniques has grown in numerous research areas. Currently, it is also widely used in statistics, including the official statistics for data collection (e.g. satellite imagery, web scraping and text mining, data cleaning, integration and imputation) but also for data analysis. However, the usage of these methods in survey sampling including small area estimation is still very limited. Therefore, we propose a predictor supported by these algorithms which can be used to predict any population or subpopulation characteristics based on cross-sectional and longitudinal data. Machine learning methods have already been shown to be very powerful in identifying and modelling complex and nonlinear relationships between the variables, which means that they have very good properties in case of strong departures from the classic assumptions. Therefore, we analyse the performance of our proposal under a different set-up, in our opinion of greater importance in real-life surveys. We study only small departures from the assumed model, to show that our proposal is a good alternative in this case as well, even in comparison with optimal methods under the model. What is more, we propose the method of the accuracy estimation of machine learning predictors, giving the possibility of the accuracy comparison with classic methods, where the accuracy is measured as in survey sampling practice. The solution of this problem is indicated in the literature as one of the key issues in integration of these approaches. The simulation studies are based on a real, longitudinal dataset, freely available from the Polish Local Data Bank, where the prediction problem of subpopulation characteristics in the last period, with "borrowing strength" from other subpopulations and time periods, is considered.