A new paper from Google DeepMind and Princeton researchers provides causal evidence that Large Language Models (LLMs) do not just passively report confidence, but actively use it to guide their behavior. The study demonstrates that by manipulating the confidence levels of these models, researchers can influence whether the LLMs choose to answer a question or abstain from answering. AI
IMPACT Demonstrates LLMs actively use confidence scores, suggesting potential for more nuanced AI behavior and control mechanisms.
RANK_REASON The cluster contains a research paper published in a scientific journal. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
