Improving Decision Models with Constrained Outputs

The article explores building decision models that constrain outputs to a fixed set of options, improving efficiency and accuracy. Traditional models require multiple passes to generate valid responses, but constrained models can select the highest probability output in a single pass. The author demonstrates this using a language model (Qwen/Qwen3-1.7B) and a simple example, showing how the model can make reasonable predictions. Testing against the CommonsenseQA dataset reveals an accuracy of about 59%, which improves slightly with finetuning. The article also highlights the issue of model overconfidence, particularly in ambiguous scenarios, and suggests temperature scaling as a method to calibrate the model’s confidence scores to better reflect its accuracy.