New Research Questions Correctness of AI Demonstrations in Learning

A new research paper explores the counterintuitive phenomenon in in-context learning (ICL) where correct demonstrations can sometimes decrease accuracy. The study introduces task-preserving perturbations to analyze this "correctness-utility gap," demonstrating that changes in exemplar input can alter the model's evidence mixture, leading to degraded performance. This effect is particularly pronounced in smaller models and more challenging tasks. AI

RANK_REASON Academic paper published on arXiv detailing a new finding in AI research. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• arXiv cs.AI
• In-context learning

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →