Technology#02 in September 2026's curated selectionCausal evidence that language models use confidence to drive behaviour
Dharshan Kumaran, Nathaniel Daw, Simon Osindero et al.
OPENSTEM EDITORIAL SUMMARY
What the research examines
This study investigates how confidence-related signals affect language-model decisions about when to answer or abstain. It provides experimental evidence that changes to such signals can influence model behaviour.
Why it may matter
Model calibration, selective answering and AI reliability.
What to examine carefully
The reported causal findings are specific to the models and experimental settings studied.
Editorial scope: This is a concise, independently written introduction based on the publisher information, not a full systematic review or independent replication. It must not replace reading the original methods, limitations and conclusions.
STUDENT RESEARCH LENS · ORIGINAL OPENSTEM NOTES
Try a small, reproducible follow-up
Design a small, openly documented benchmark that measures answer accuracy, confidence and abstention on questions with known answers. Use a model you are authorized to evaluate.
Questions to investigate
- How does the paper identify confidence-related signals and intervene on them?
- Are changes in abstention accompanied by changes in actual answer correctness?
These are independently proposed educational investigations, not activities that the original research authors claim to have completed.
Publication & researcher information
JournalNature Machine Intelligence
Published online2026-09-07
Authors (selected)Dharshan Kumaran, Nathaniel Daw, Simon Osindero et al.
University / laboratorySee original article for verified affiliations
CountryNot independently verified for all coauthors
Corresponding authorConsult publisher correspondence section
Multi-author research may involve universities in several countries. Unverified institutions and contact addresses are not inferred from authors' names.