In human cognition theory, human thinking is governed by two systems: the fast and intuitive System 1 and the slower but more deliberative System 2. Recent studies have shown that incorporating System 2 process into Transformers including large language models (LLMs), significantly enhances their reasoning capabilities. Nevertheless, models that purely resemble System 2 thinking require substantially higher computational costs and are much slower to respond. To address this challenge, we present Dualformer, a single Transformer model that seamlessly integrates both the fast and slow reasoning modes. Dualformer is obtained by training on data with randomized reasoning traces, where different parts of the traces are dropped during training. The dropping strategies are specifically tailored according to the trace structure, analogous to analyzing our thinking process and creating shortcuts with patterns. At inference time, our model can be configured to output only the solutions (fast mode) or both the reasoning chain and the final solution (slow mode), or automatically decide which mode to engage (auto mode). In all cases, Dualformer outperforms the corresponding baseline models in both performance and computational efficiency: (1) in slow mode, Dualformer optimally solves unseen 30 x 30 maze navigation tasks 97.6% of the time, surpassing the Searchformer (trained on data with complete reasoning traces) baseline performance of 93.3%, while only using 45.5% fewer reasoning steps; (2) in fast mode, Dualformer completes those tasks with an 80% optimal rate, significantly outperforming the Solution-Only model (trained on solution-only data), which has an optimal rate of only 30%. For math problems, our techniques have also achieved improved performance with LLM fine-tuning, showing its generalization beyond task-specific models.

While transformers have shown great potential on video recognition tasks with their strong capability of capturing long-range dependencies, they often suffer high computational costs induced by self-attention operation on the huge number of 3D tokens in a video. In this paper, we propose a new transformer architecture, termed DualFormer, which can effectively and efficiently perform space-time attention for video recognition. Specifically, our DualFormer stratifies the full space-time attention into dual cascaded levels, i.e., to first learn fine-grained local space-time interactions among nearby 3D tokens, followed by the capture of coarse-grained global dependencies between the query token and the coarse-grained global pyramid contexts. Different from existing methods that apply space-time factorization or restrict attention computations within local windows for improving efficiency, our local-global stratified strategy can well capture both short- and long-range spatiotemporal dependencies, and meanwhile greatly reduces the number of keys and values in attention computation to boost efficiency. Experimental results show the superiority of DualFormer on five video benchmarks against existing methods. In particular, DualFormer sets new state-of-the-art 82.9%/85.2% top-1 accuracy on Kinetics-400/600 with around 1000G inference FLOPs which is at least 3.2 times fewer than existing methods with similar performances.