Placed in an examination situation under common time constraints, they behave like gluttonous sequential solvers: they exhaust themselves on the first questions and ignore their value.
What happens when a reasoning model has to deal with multiple problems under a single constraint of cost or latency? The team of Chenrui Fan, Yize Cheng, Ming Li, Yongyuan Liang, Tianyi Zhou, and Soheil Feizi asked the question in a paper filed on August 8, 2026, and the answer is unflattering: models don't strategically allocate this shared budget.
Their behavior is similar to that of a gluttonous sequential solver. They deal with the questions in the order in which they are presented, concentrate most of the effort on the first ones, and remain insensitive to the value assigned to each one. The phenomenon worsens as the number of questions increases—exactly the opposite of what would be required to manage scarce resources properly.
The authors tested explicit planning prompts: they improve the distribution, without producing a prioritization that takes into account the difficulty or stakes of each question. The pattern can be observed on both mathematical tasks and code reasoning, which suggests the absence, in current models, of an overall budget arbitrage capacity rather than a weakness specific to a domain.
Read the full article on the original source
Go to the source →