A new research paper explores how different types of language models and humans process repeated words. Base large language models (LLMs) exhibit automatic processing, showing consistent facilitation regardless of lag or context removal. In contrast, instruction-tuned LLMs demonstrate controlled processing, with facilitation decaying over time and even reversing to interference at larger scales. The study found that within the Qwen 2.5 family, larger models increasingly alter repetition processing. While humans show a hybrid profile, neither LLM type fully replicates human cognitive processes. AI
IMPACT Reveals fundamental differences in how LLMs process language compared to humans, suggesting post-training significantly alters model behavior.
RANK_REASON The cluster contains a research paper detailing findings on LLM processing. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
