LLM Self-Report Tracks Internal Activations скачать в хорошем качестве

LLM Self-Report Tracks Internal Activations

In this AI Research Roundup episode, Alex discusses the paper: 'When Models Examine Themselves: Vocabulary-Activation Correspondence in Self-Referential Processing' This study investigates whether Large Language Models are actually reporting internal computation or just confabulating when prompted for self-examination. Using the new Pull Methodology, researchers discovered that self-referential vocabulary in models like Llama 3.1 and Qwen 2.5 directly tracks specific internal activation dynamics. For example, using loop vocabulary correlates with higher activation autocorrelation, while shimmer vocabulary tracks increased activation variability. This correspondence is unique to self-referential contexts and suggests that model self-reports can reliably reflect internal computational states. The findings reveal a specific activation direction for self-referential processing that is distinct from refusal or descriptive tasks. Paper URL: https://arxiv.org/pdf/2602.11358 #AI #MachineLearning #DeepLearning #LLM #Interpretability #ModelIntrospection #Llama3 #TransformerModels