The BasePeople

Maria Victoria Carro

Università degli Studi di Genova; FAIR, IALAB, Universidad de Buenos Aires

Titles as the sources gave them. In 1 entry, with 4 checked quotations.

What they said

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

“For weaker language models, however, the results are mixed: debate provides no comparable gains across most settings, with effects ranging from significant improvements to declines in accuracy.”
Humans Correct Their Judgments Through Debate, but Weaker AI Models Do NotarXiv, October 20, 2026. Speaking as: Other. Summary of the LLM-judge experiments in the introduction, after the authors report accuracy gains for human judges.
“For LLM supervisors, however, none of them significantly improves group accuracy, even as the number of judges increases.”
Humans Correct Their Judgments Through Debate, but Weaker AI Models Do NotarXiv, October 20, 2026. Speaking as: Other. Result on multi-judge protocols (voting, deliberation) with language-model judges, contrasted with human groups.
“Notably, in non-diverse deliberation panels, grounded minority judges abandon their initial answers more often than erroneous minority judges adopt the correct answer, even when the grounded minority judges report greater confidence in their pre-deliberation positions.”
Humans Correct Their Judgments Through Debate, but Weaker AI Models Do NotarXiv, October 20, 2026. Speaking as: Other. The authors' observation about deliberation among identical judges, where they say incorrect positions may be more persuasive.
“Taken together, these findings provide evidence supporting Debate as a viable path toward scalable human oversight, while casting doubt on the assumption that such benefits translate to weaker AI supervisors, raising concerns about the reliability of fully automated training and control pipelines in high-stakes AI safety contexts.”
Humans Correct Their Judgments Through Debate, but Weaker AI Models Do NotarXiv, October 20, 2026. Speaking as: Other. The authors' overall reading of their results at the end of the introduction.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

    Humans Correct Their Judgments Through Debate, but Weaker AI Models Do Not

    A team led by Maria Victoria Carro tests AI Safety via Debate, in which two AI debaters argue opposite answers before a judge. With 507 human participants and over 40 LLM experiments, they report debate improved human accuracy but gave weaker language-model judges little consistent gain.