The problem: closed-choice decisions between speed and accuracy
Many uses of language models end in a closed-choice decision: one option from a defined set, whose probabilities can be used by the code. According to the study published on arXiv, reading these probabilities in a single forward pass is fast but often errs on tasks requiring multiple steps, such as arithmetic, counting and date comparison. Making the model reason first corrects most of these errors, but costs seconds per decision instead of milliseconds.
The work, titled ThinkJev, therefore addresses a practical trade-off: spending additional compute only where it produces a measurable benefit, without giving up the typed output format that code can act on.
How it works: two systems from a single hybrid model
ThinkJev builds a fast system and a slow system from a single open model with hybrid reasoning capabilities, able to answer directly or after a thinking pass. The first system reads the probabilities associated with the option letters when the model answers immediately; the second reads the same letters after the model has reasoned over the same cached prompt. Because the escalation restarts from the cache, according to the authors it does not repeat work, and the two systems produce probabilities over the same options.
A central element is that the letters can be read after any number of thinking tokens: a single sampled thought thus provides the answer under every reasoning budget. The system can also stop thinking as soon as the answer becomes sufficiently certain.
The reported numbers and the stated limitations
The fast system answers on its own only where a statistical test on labeled development data guarantees that accepted answers remain below a predefined error threshold, without training. With a 5% target, Qwen3.5-27B answers this way on 24.5% of 8,359 test decisions, with a realized error of 3.9%; for Qwen3.5-4B and 9B the reported values are 6.5% and 0.0%. For the remaining cases, stopping thinking as soon as the answer is certain reaches 81.6% accuracy, versus 82.7% for always reasoning, using 72% of the thinking tokens.
The authors also report a negative result: choosing when to think based only on the fast system's confidence saves less, because the decisions it is least sure about are often not those that reasoning corrects. On the public JevBench questions, an independent benchmark for typed decision services, Qwen3.5-27B reasons in 29.4% of cases and is not distinguishable in accuracy from the two listed reasoning APIs; the authors qualify this comparison as exploratory. Code and data generators are said to be available.
What remains to be verified
The results concern specific configurations and models and a benchmark described as exploratory in the comparison with external services; the material does not clarify behavior on other domains or overall production costs. The source is the article's arXiv page, which contains no independent reactions.
