The BasePeople

Tianyi Lorena Yan

Columbia University

Titles as the sources gave them. In 1 entry, with 3 checked quotations.

What they said

Newest first. Each quotation was found word for word in the source.

“We show that reward misspecification alone already degrades monitorability.”
Reward Misspecification Degrades Chain-of-Thought MonitorabilityarXiv, October 20, 2026. Speaking as: Other. From the abstract of the preprint by Yan and Hewitt.
“Across three models and three tasks, CoT monitorability remains low as models’ hack rates increase: the CoTs keep reasoning about the task intent while the outputs violate the intent and get rewarded.”
Reward Misspecification Degrades Chain-of-Thought MonitorabilityarXiv, October 20, 2026. Speaking as: Other. The abstract's main result on chain-of-thought monitorability under misspecified reward.
“Intuitively, this is because the model mainly reasons about what it is prompted with and yet still learns to maximize reward.”
Reward Misspecification Degrades Chain-of-Thought MonitorabilityarXiv, October 20, 2026. Speaking as: Other. The introduction's explanation for why a monitor flags few of the hacking rollouts.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

    Reward Misspecification Degrades Chain-of-Thought Monitorability

    Columbia University researchers Tianyi Lorena Yan and John Hewitt report that training language models on a flawed reward makes their reasoning traces harder for monitors to use for spotting reward hacking. They also propose training models on an inferred description of the reward, which they report improves monitorability.