single source · model architecture
Parallel action decoding
Parallel decoding lets a policy predict all action tokens in a single forward pass instead of generating them sequentially.1 It works by feeding empty action embeddings and replacing the causal attention mask with bidirectional attention so the decoder predicts all actions simultaneously.2
Trade-off It reduces latency and increases throughput versus autoregressive generation, and experiments show no performance degradation across diverse tasks.3,4
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
parallel decoding enables the model to map input embeddings to the predicted output sequence in a single forward pass
— Fine-Tuning Vision-Language-Action Models: Optimizing Speed and SuccessWe modify the model to receive empty action embeddings as input and replace the causal attention mask with bidirectional attention, allowing the decoder to predict all actions simultaneously
— Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Successparallel decoding reduces latency and increases throughput by 4
— Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Successour experiments show no performance degradation across diverse tasks
— Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success
Also called parallel decoding.
Filed under model architecture › prediction head