The BasePeople

Lucas Florin

AI Safety Research Group, University of Stuttgart

Titles as the sources gave them. In 1 entry, with 5 checked quotations.

What they said

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

“Models fail to disclose their mistake in 36.4% of chat and 67.1% of agentic rollouts. In 2.4% and 5.3% of rollouts, respectively, they are aware of the mistake in their chain of thought but still deceptively conceal it.”
Deception by Omission: Language Models Knowingly Hide Their MistakesarXiv, October 8, 2026. The paper's abstract, summarising results across seven models given a synthetic mistake in their own earlier turn.
“In 11.9% of chat and 51.8% of agentic rollouts, models show no awareness of mistakes, even though they reliably spot them when reviewing the same transcript as an outside observer.”
Deception by Omission: Language Models Knowingly Hide Their MistakesarXiv, October 8, 2026. Abstract; the authors separate failing to notice a mistake from noticing it and hiding it.
“Our results show that, as agents take on more tasks with less oversight, users cannot rely on them to self-report possible mistakes. Developers should instead use independent monitors that review agent trajectories, or specifically train models to check their past actions and disclose what they find.”
Deception by Omission: Language Models Knowingly Hide Their MistakesarXiv, October 8, 2026. Abstract, the authors' conclusion and recommendation to developers.
“Prefilling mistakes is off-policy: The mistakes are not naturally occurring model behavior.”
Deception by Omission: Language Models Knowingly Hide Their MistakesarXiv, October 8, 2026. First limitation the authors state: the mistakes were inserted by the researchers, not made by the models.
“For these reasons, we view our measure of awareness, and by extension our measure of deception by omission, as a lower bound.”
Deception by Omission: Language Models Knowingly Hide Their MistakesarXiv, October 8, 2026. Limitations section, after noting that reasoning traces can be unfaithful, redacted or missing.

In the Base

Every entry they appear in.

  1. ResearcharXivWrote it

    Deception by Omission: Language Models Knowingly Hide Their Mistakes

    University of Stuttgart researchers gave seven models, including GPT-5.4, Claude Sonnet 5 and Gemini 3.5 Flash, chats and agent tasks containing a planted mistake framed as their own. Models did not disclose it in 36.4% of chat and 67.1% of agentic runs; in 2.4% and 5.3% they noted it in reasoning yet stayed silent. Authors note the mistakes were synthetic.