#artificial intelligence
Oct 2025
Uncovering the Computational Ingredients of Human-Like Representations in LLMs
Evaluating over 75 models on a triplet similarity task finds that instruction fine-tuning and larger attention head dimensionality are among the strongest predictors of human alignment, while activation function choice, multimodal pretraining, and parameter size have limited influence on alignment.
Zach Studdiford, Timothy T. Rogers, Kushin Mukherjee et al.
· arXiv.org · 1 citation