New training lifts math answers at 5% risk from 0.8% to 42.4%
CARE rewards a confidence score by the scores for other questions, not by the answer only.
Researchers made CARE, a method to train a language model to tell how sure it is of an answer. A system can then abstain, or give the question to a person or a stronger model, when the model is not sure. On math tests, the share of questions that the model answers at 5% risk of error rose from 0.8% to 42.4%. The results are compared with older rewards that score each answer only.
How it works
Older methods give a reward of 1 for a correct answer and 0 for an incorrect answer. Scores pushed to 0 and 1 do not show how answers compare with each other. CARE looks at a set of answers. At each risk level, the scores set a threshold, and each score gets a reward for the correct side of it.
The numbers
The math tests are MATH-500, GSM8K and Big-Math-Digits. The team also used multi-hop Wikipedia questions for training. On tests external to this training, such as GPQA and TriviaQA, the share rose from 12.3% to 23.6%.
What is not known
The abstract gives no data on cost of training. It does not tell how the method works in use. All results are from the authors.
Insight: Whoever sets the reward sets how honest the confidence is
In October 2026, researchers found that a model gave high confidence on 56% of tasks with a high risk of failure. The researchers find that incentives cause this. CARE changes the incentive for the score. The confidence of a model follows its reward, and the vendor that trains it sets how far you can trust the score. For leaders, find out how a vendor trains confidence.
Sources
Posted