This paper highlights the limitations of the "multiple-input, single-output" (MISO) architecture employed by existing large-scale pre-trained models, such as ChatGPT and OpenVLA. This architecture causes task mutual exclusion in "multiple-input, multiple-output" (MIMO) tasks (e.g., parallel multi-task output processing), leading to resource competition among multiple tasks when sharing output channels, resulting in optimization imbalances and performance degradation. In contrast, humans can simultaneously execute tasks without interference through MIMO processing (e.g., concurrent conversation and decision-making). Inspired by this, we propose the Visual Language Action Model for Simultaneously Chatting and Decision Making (VLASCD, or MIMO-VLA), an integrated MIMO-trained model with parallel multi-task outputs capable of simultaneous conversation and decision-making. Experimental results on the CARLA autonomous driving platform demonstrate that MIMO-VLA significantly outperforms LLM models with MISO conversation capabilities, reinforcement learning models, and VLA models with MISO decision capabilities in simultaneously processing conversation and decision-making tasks in MIMO scenarios.
