<ul><li>A single Transformer model called Dualformer is presented, which integrates both fast and slow reasoning modes.</li><li>Dualformer is trained on data with randomized reasoning traces, dropping different parts of the traces during training.</li><li>At inference time, Dualformer can be configured to output only solutions (fast mode), reasoning chain and solution (slow mode), or automatically decide which mode to engage (auto mode).</li><li>In terms of performance and computational efficiency, Dualformer outperforms corresponding baseline models, showing improved performance in maze navigation tasks and math problems.</li></ul>

Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces

Discover more