LLM factual recall scales with model size and training data frequency

Researchers have identified a predictable relationship between factual recall in large language models, their size, and the frequency of topics in their training data. By evaluating 38 models on over 8,900 scholarly references, they found that recall quality follows a sigmoid curve based on a combination of model parameters and topic representation. These factors alone accounted for a significant portion of the variance in recall performance across different model families. AI

IMPACT Establishes a new scaling law for factual recall in LLMs, suggesting that performance is predictable based on model size and training data composition.

RANK_REASON Academic paper detailing a new finding about LLM behavior. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• LLMs
• training data
• model size

• paper
• model release

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