Large language models frequently generate lengthy reasoning traces even for simple decisions, driving up computational cost. A common remedy is to impose a reasoning budget, but this can force the model to answer from an incomplete chain of thought, which may degrade accuracy. The tension between saving tokens and preserving correctness is the central problem the paper addresses.
The authors propose evaluating System-1 decision models, which are designed to select among options directly rather than through extended deliberation. By comparing these models against standard LLMs under truncated reasoning budgets, the study aims to determine whether faster, more economical decision-making can match the accuracy of full reasoning.
Because the abstract is the only source here, specific results are not yet available. The paper's significance lies in framing the trade-off and suggesting that System-1 models could be a practical alternative when budgets are tight. Further details would require reading the full preprint.