A technical evaluation of constrained decoding methods shows that language models often assign inflated confidence to incorrect multiple-choice answers. Researchers tested a 1.7-billion-parameter model on structured prediction tasks where outputs were limited to predefined options such as A, B, C, D, or E. When presented with questions requiring single-token responses, the model generated predictions after multiple inference passes, with each pass corresponding to a token in the output sequence.

In one test, the model correctly identified the color of the sky as blue when constrained to the options, assigning a probability of 0.9988 to that choice. However, in other cases, the model demonstrated high confidence in answers that were statistically no more likely to be correct than random selections. For example, when asked where a bat would most likely be found, the model predicted "Cave" with 0.9978 probability despite the absence of a clearly correct answer in the context. Calibration tests revealed that the model's confidence scores did not align with actual accuracy rates across different probability ranges.

The study found that the model tended to be most overconfident in predictions where the assigned probability fell between 0.90 and 1.00, a range where its accuracy was only about 70%. Even when confidence was moderate, such as in the 0.70 to 0.80 range, accuracy improved only marginally. These discrepancies suggest that raw probability outputs from such models cannot be reliably interpreted as measures of correctness without further calibration.

To address this, researchers applied temperature scaling, adjusting the softmax function's temperature parameter to flatten probability distributions and better align confidence with empirical performance. Testing indicated that a temperature value of approximately 3.797 improved calibration, bringing predicted accuracy closer to observed results across confidence bins. The findings highlight limitations in using constrained decoding alone for high-stakes decision-making and underscore the need for additional validation when relying on model outputs in structured settings.