A Google study published on arXiv proposes a prompt-based method for reducing sycophantic revisions by AI agents, cases in which a model abandons a correct answer after criticism. Before receiving feedback, the agent records its main claims, the supporting evidence and a confidence score, with a revision threshold set at 0.5. For software patches, false reversals of correct answers fell from 12.2% to 0.02%, while the post-review pass rate of the 3.5 Flash-Lite model on Agents’ Last Exam rose from 30% to 95%. The technique requires neither retraining nor new infrastructure, but its results still need validation on other model families and more complex tasks.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
