Upgrade the model of the study of the study chain-of-thought to think
Chain-of-tempent (CON) Reasoning in Vision Language Models (VLMS) is essential to improving interpretation and trust. However, the current methods of such relying training often depends on the datasets with short adjectives with smaller measurements. In this work, we show that short-refund training leads to complete measurement in consultation activities that require more detailed explanations. Dealing with this limit, we raise a strategy with two training stage that increase the use of short response details to develop the Cot Preming. First, we add short answers to GPT-4O COT consultation, enhancing CLM's cot power in good order. Second, we find short answers as rewards of strengthening. Specially, short answers are used as indicators of formalization of bad papers (right) that is not good from the chains of the consultation produced. These two are used to estimate the model thinking about the highest Direct. Our testing shows a major development in the Canchmark Datasets, as well as regular development to direct response predictions. This work provides for the resources of the Vlm Cot training resources and shows the effective performance of the multimodal models post – post-multimodal training.
- † work done while in apple
- ‡ Conongie Mellon University
