Monitoring Monitorability

Melody Y. Guan, Miles Wang, Micah Carroll, Zehao Dou, Annie Y. Wei, Marcus Williams, Benjamin Arnav, Joost Huizinga, Ian D Kivlichan, Amelia Glaese, Jakub Pachocki, Bowen Baker
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:37538-37602, 2026.

Abstract

Safe deployment of increasingly capable AI agents may require visibility into how they make decisions. Chain-of-thought (CoT) monitoring can detect misbehavior in today’s reasoning models, but this “monitorability” may be fragile under different training procedures, data sources, or continued system scaling. We propose three evaluation archetypes (intervention, process, and outcome-property), a new monitorability metric, and a broad evaluation suite. We show CoT monitoring outperforms action-only monitoring in practical settings, and that frontier models are generally—but not perfectly—monitorable. We study scaling trends with pre-training model size and inference-time compute, finding longer CoTs are typically more monitorable. We find that, for a fixed capability level, using a smaller model at higher reasoning effort can yield higher monitorability, at greater inference compute cost. We further find that increasing a weak monitor’s test-time compute when monitoring a strong agent improves monitorability, and giving the monitor access to the CoT both boosts monitorability and steepens the compute–to-monitorability scaling trend. Finally, we show monitorability can be improved by asking follow-up questions and giving the follow-up CoT to the monitor.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-guan26h, title = {Monitoring Monitorability}, author = {Guan, Melody Y. and Wang, Miles and Carroll, Micah and Dou, Zehao and Wei, Annie Y. and Williams, Marcus and Arnav, Benjamin and Huizinga, Joost and Kivlichan, Ian D and Glaese, Amelia and Pachocki, Jakub and Baker, Bowen}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {37538--37602}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/guan26h/guan26h.pdf}, url = {https://proceedings.mlr.press/v306/guan26h.html}, abstract = {Safe deployment of increasingly capable AI agents may require visibility into how they make decisions. Chain-of-thought (CoT) monitoring can detect misbehavior in today’s reasoning models, but this “monitorability” may be fragile under different training procedures, data sources, or continued system scaling. We propose three evaluation archetypes (intervention, process, and outcome-property), a new monitorability metric, and a broad evaluation suite. We show CoT monitoring outperforms action-only monitoring in practical settings, and that frontier models are generally—but not perfectly—monitorable. We study scaling trends with pre-training model size and inference-time compute, finding longer CoTs are typically more monitorable. We find that, for a fixed capability level, using a smaller model at higher reasoning effort can yield higher monitorability, at greater inference compute cost. We further find that increasing a weak monitor’s test-time compute when monitoring a strong agent improves monitorability, and giving the monitor access to the CoT both boosts monitorability and steepens the compute–to-monitorability scaling trend. Finally, we show monitorability can be improved by asking follow-up questions and giving the follow-up CoT to the monitor.} }
Endnote
%0 Conference Paper %T Monitoring Monitorability %A Melody Y. Guan %A Miles Wang %A Micah Carroll %A Zehao Dou %A Annie Y. Wei %A Marcus Williams %A Benjamin Arnav %A Joost Huizinga %A Ian D Kivlichan %A Amelia Glaese %A Jakub Pachocki %A Bowen Baker %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-guan26h %I PMLR %P 37538--37602 %U https://proceedings.mlr.press/v306/guan26h.html %V 306 %X Safe deployment of increasingly capable AI agents may require visibility into how they make decisions. Chain-of-thought (CoT) monitoring can detect misbehavior in today’s reasoning models, but this “monitorability” may be fragile under different training procedures, data sources, or continued system scaling. We propose three evaluation archetypes (intervention, process, and outcome-property), a new monitorability metric, and a broad evaluation suite. We show CoT monitoring outperforms action-only monitoring in practical settings, and that frontier models are generally—but not perfectly—monitorable. We study scaling trends with pre-training model size and inference-time compute, finding longer CoTs are typically more monitorable. We find that, for a fixed capability level, using a smaller model at higher reasoning effort can yield higher monitorability, at greater inference compute cost. We further find that increasing a weak monitor’s test-time compute when monitoring a strong agent improves monitorability, and giving the monitor access to the CoT both boosts monitorability and steepens the compute–to-monitorability scaling trend. Finally, we show monitorability can be improved by asking follow-up questions and giving the follow-up CoT to the monitor.
APA
Guan, M.Y., Wang, M., Carroll, M., Dou, Z., Wei, A.Y., Williams, M., Arnav, B., Huizinga, J., Kivlichan, I.D., Glaese, A., Pachocki, J. & Baker, B.. (2026). Monitoring Monitorability. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:37538-37602 Available from https://proceedings.mlr.press/v306/guan26h.html.

Related Material