A new research paper explores the inner workings of In-context Learning (ICL) in large language models, proposing that ICL functions by selectively removing task-irrelevant information from the model's internal representations. The study suggests that in a zero-shot scenario, models produce arbitrary outputs because their representations contain information for all tasks. However, few-shot demonstrations effectively guide the model to discard redundant information, thereby focusing on the intended task. Researchers identified specific 'Denoising Heads' within the model's attention mechanisms that are crucial for this information removal process, demonstrating that disabling these heads significantly degrades ICL accuracy. AI
IMPACT Provides a new theoretical framework for understanding and potentially improving In-context Learning capabilities in LLMs.
RANK_REASON Research paper detailing a novel mechanism for In-context Learning in Language Models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
