Fine-tuning four reasoning models to predict their own confidence scores reduced their average generated tokens by 10.3% to 19.2% against the corresponding base models across five math, science and coding benchmarks. The paper reports statistically significant token reductions, while accuracy changes were not statistically significant.
At selected points in a reasoning trace, each model produced a tentative answer. ConfSFT, the researchers’ method, combined the model’s probabilities for the answer tokens into a score adjusted for answer length. It then trained the model to predict that score from its reasoning so far. The label needed neither a gold answer nor an outside judge, and it is not a calibrated probability that the answer is correct.
Across 480 Nemotron traces on AIME 2024, higher scores were associated with tentative answers more likely to be correct and to match the model’s eventual final answer. The average accuracy gain from continuing the trace also fell as confidence rose. The authors use these associations to motivate confidence as a training signal; they do not establish that it caused the shorter outputs.
The evaluation covered Nemotron-Nano-8B, Gemma-4-E2B, Qwen3-4B and GPT-OSS-20B. Average token reductions were 11.1%, 10.3%, 19.2% and 11.1%, respectively, measured against each model’s base version across AIME 2025, GSM8K, GPQA-Diamond, HumanEval and LiveCodeBench. Training used AIME problems from 2000–2023, with AIME 2024 held out for validation; each benchmark problem received 16 sampled completions.
Figure 5 compares changes in token share across eight reasoning categories. ConfSFT shifted the models’ category mix less than several tested efficiency methods, according to the paper. At inference, the models used ordinary decoding without a confidence prompt or early-stop rule. The study measured token counts, not response time, energy use or deployed serving cost, so those operational effects remain untested.