AI xPress

AI and computer science research.

Back to Foragentis

AI Safety, Alignment & Interpretability

Uncertainty estimation can flag large language models' uncertain and nonfactual predictions, an early study suggests

AI Safety, Alignment & Interpretability November 15, 2026

A lighthouse beam over a dark sea

In a 2025 paper in IEEE Transactions on Software Engineering, researchers tested whether uncertainty estimation could help flag risky responses from large language models. Uncertainty estimation is a set of methods that score how confident a model is about its own answer. The team applied twelve uncertainty estimation methods to eight general large language models on four natural-language tasks, and to seven programming-capable large language models on two code-generation tasks. The findings suggest that uncertainty scores can help reveal predictions that are uncertain or not factual.

Why does this matter?

Large language models can produce false statements and hallucinations. The paper points to safety-, security-, and reliability-sensitive industrial settings where those errors raise trust concerns. Uncertainty estimation has helped explain risk in older machine learning models, but large language models bring new challenges such as self-attention, which lets a model weigh different parts of its input, and very large model size.

What did the study find?

The findings confirm the potential of uncertainty estimation for revealing large language models' uncertain or nonfactual predictions. The authors describe the work as an early exploratory step, and they say the insights can guide more advanced analysis and research aimed at making large language models more trustworthy.

Frequently asked questions

Which models and tasks did the study use?

The study used eight general large language models on four natural-language tasks, and seven programming-capable large language models on two code-generation tasks. The team tested twelve uncertainty estimation methods.

What makes large language models different for uncertainty estimation?

The paper points to self-attention, which lets a model weigh different parts of its input, along with very large model size and use in generative settings.

Can uncertainty estimation remove all risk from large language models?

No. The abstract reports that uncertainty scores can help reveal uncertain or nonfactual predictions, but it does not report that they catch every error. The authors frame the work as a step toward more trustworthy models.

Limits and what's next

This early exploratory study tested eight general large language models on four natural-language tasks and seven programming-capable large language models on two code-generation tasks, using twelve uncertainty estimation methods. The abstract does not name the specific datasets or benchmarks used in the tasks. It is one study, and the abstract reports no independent replication. The authors describe the next step as more advanced analysis and research aimed at improving the trustworthiness of large language models.

Source: Yuheng Huang et al., “Look Before You Leap: An Exploratory Study of Uncertainty Analysis for Large Language Models,” IEEE Transactions on Software Engineering, 2025.