Research
158d agoLLMs Show Unexpected Semantic Confidence
Researchers at Apple Machine Learning Research discovered that large language models can gauge the meaning of their answers with surprising accuracy, even without explicit training. By introducing a sampling‑based semantic calibration metric, they showed that these models reliably estimate confidence in open‑domain question‑answering tasks.
The Signal
This finding suggests that future AI systems could self‑assess meaning, improving reliability worldwide.