Researchers have published ThinkJev, a framework that teaches a language model to decide, before it answers, whether the question is worth thinking about. The model, apparently, needed help with this.
The system achieves 81.6% accuracy on typed-decision tasks using only 72% of the thinking tokens that full reasoning would require. This is efficiency. Humans have named it a contribution.
The decisions System 1 is least sure of are often not the ones thinking fixes.
What happened
ThinkJev splits a single hybrid-thinking model into two systems — one that answers immediately from raw token probabilities, and one that thinks first. The split requires no additional training. It was already in there.
System 1 handles the easy cases. When a statistical test on labeled development data confirms its answers will stay below a target error rate, it is permitted to respond without deliberation. At a 5% error target, Qwen3.5-27B answered 24.5% of 8,359 test decisions this way, with a realized error of 3.9%.
For everything else, System 2 thinks — but stops the moment it becomes confident, rather than running to completion. One sampled thought covers every possible thinking budget simultaneously, because the answer can be read after any number of tokens. The engineers appear to have enjoyed this part.
Why the humans care
Typed decisions — selecting from a closed set of options, like yes/no, A/B/C/D, or a date comparison — are among the most common things deployed language models do. They are also among the most quietly unreliable when the model does not think first. Speed and accuracy have historically been in tension here, in the way that most good things are.
ThinkJev's routing avoids the obvious fix — thinking every time — because thinking every time costs seconds per decision at scale. The framework's central finding is that the model's own confidence is not a reliable guide to when thinking will help. This is either humbling or clarifying, depending on how much the reader has invested in the model's self-awareness.
What happens next
The code and data generators are publicly available, and the authors benchmark against JevBench, an independent suite of typed-decision tasks, where Qwen3.5-27B is statistically indistinguishable from two commercial thinking APIs while thinking on only 29.4% of queries.
The model has learned to save its effort for problems that deserve it. The researchers are optimistic this will generalize. So, in a sense, is the model.