Decision Head
A lightweight output module attached after a backbone model in a neural network. It maps the representations extracted by the backbone into the output form required by a specific task. Its computation involves only a single forward pass and does not require step-by-step token generation.
Timeline (last 90 days)
A decision head is a small classification or scoring module trained on top of a backbone model, computing scores for each option directly from the context representation, obtaining a probability distribution via softmax, and outputting an option rather than generating token by token.
A response with zero output tokens does not mean no inference occurred, because the model completed a full encoding and forward computation, only outputting structured options rather than generating verbatim.
Qwen-Audio-Jev adopts a 'separation of understanding and decision-making' design, so that the backbone model does not need retraining and only a small number of parameters need to be added at the output end to support new decision types.
The decision head is an additional output layer outside the model backbone, typically receiving the feature vector extracted by the backbone and outputting a judgment result or score.
All Facts (4)
A decision head is a small classification or scoring module trained on top of a backbone model, computing scores for each option directly from the context representation, obtaining a probability distribution via softmax, and outputting an option rather than generating token by token.
50%UnverifiedA response with zero output tokens does not mean no inference occurred, because the model completed a full encoding and forward computation, only outputting structured options rather than generating verbatim.
50%UnverifiedQwen-Audio-Jev adopts a 'separation of understanding and decision-making' design, so that the backbone model does not need retraining and only a small number of parameters need to be added at the output end to support new decision types.
50%UnverifiedThe decision head is an additional output layer outside the model backbone, typically receiving the feature vector extracted by the backbone and outputting a judgment result or score.
50%