Repetition Priming Study Reveals Divergent Processing in Base and Instruct LLMs
A recent investigation published on arXiv (2608.14681) examines the way large language models (LLMs) handle repeated words, contrasting base models, instruct models, and human responses. Utilizing the psychological concept of repetition priming, the study evaluates 15 models from five different families (ranging from 1.5B to 14B parameters) through semantic categorization and cloze completion tasks, alongside comparable human experiments. Results indicate that base models demonstrate automatic processing, showing stable immediate facilitation across different lags, which partially persists even when context is removed and aligns with attention to previous instances. Conversely, instruct models reveal controlled processing, where facilitation diminishes with lag, fails without anticipated context, and shifts to interference at larger scales. Notably, within the Qwen 2.5 family, this dissociation escalates with model size, implying that post-training modifies inherent processing mechanisms. This research sheds light on LLMs' management of repeated words and the effects of instruction tuning on cognitive-like functions.
Key facts
- Study applies repetition priming to 15 models across five model families (1.5B-14B parameters).
- Two tasks used: semantic categorization and cloze completion.
- Matched human experiments with identical stimuli were conducted.
- Base models show automatic processing: immediate facilitation stable across lags.
- Instruct models show controlled processing: facilitation decays with lag and collapses without context.
- Instruct models reverse to interference at larger scales.
- Within Qwen 2.5 family, dissociation increases with model scale.
- Post-training changes default behavior in LLMs.
Entities
Institutions
- arXiv