A new decoding trick lets multiple-choice-only chatbots spit out exact numbers instead of vague ranges.
Researchers built NumericJev, a training-free algorithm that works on top of Jev-like language models, systems designed to answer by picking from a structured list of options rather than generating free text. Instead of asking the model to choose the closest answer from a candidate list, NumericJev has the model repeatedly narrow down a numeric range, without touching the model's internal weights or hidden states. On a 100-value grid, a ten-way branching version of this process needs just two rounds of questions to land on an answer. In testing on an arithmetic benchmark, the method's average error of 1.84 percent beat direct multiple-choice selection's 5.18 percent, and it outscored picking from a list that already contained the correct answer by 2.93 percentage points.
LLMs are notoriously bad at basic arithmetic and precise figures, and multiple-choice interfaces exist specifically to route around that weakness. NumericJev suggests you do not have to choose between structure and precision. You can keep the safety of picking from a list while still pulling out a specific number, without retraining anything or peeking inside the model's weights.
In a separate test recalling values from a historical index across three dates, the method's error crept up to 4.58 percent, a reminder that training-free does not mean solved. It just means the workaround is cheap enough to bolt onto whatever model you are already running.