Typed decision models often face a trade-off between responding quickly and reasoning more deeply. A new paper on arXiv proposes an interface that lets such models use additional reasoning when needed, without giving up the candidate outputs they have already generated. The work is titled "Selective Reasoning for Typed Decision Models: When to Start, Read, and Stop."

The authors train a shared candidate-answer reader on a frozen Qwen3.5-4B backbone. This reader is designed to unify direct decisions with on-demand reasoning, meaning the model can either answer immediately or choose to read further information before committing to a final output. The frozen backbone keeps the base model stable while the reader learns when extra reasoning is beneficial.

The paper's title highlights the decision points the method addresses: when to begin reasoning, what to read, and when to stop. This suggests a cost-aware mechanism that balances accuracy against computational effort. However, the abstract is truncated, so the full details of the training procedure and evaluation are not available from the source alone.