SourceInspection and Control of Self-Generated-Text Recognition Ability in Llama3-8b-Instruct
The authors measure the correlation between residual stream activations and the self-recognition vector at every layer. In the individual paradigm, text token activations are most correlated with the vector at layers 4-6, a pattern present in both the chat and base models and even in the absence of any prompt, suggesting a bottom-up perceptual process. A second, sharper peak appears at layer 16 at the output token and the assistant tag token. Correlations between vector activation at the output token and the model's probability of claiming self-authorship climb rapidly from layers 14-16 to a plateau, a pattern consistent across all datasets and absent in the base model or without role tags. Averaged across layers and tokens, layer-16 vector activation is higher for self-written texts (SAD: mean difference 0.0054, 95% CI [0.0045, 0.0063]; QA: 0.0061, 95% CI [0.005, 0.0073]), and this activation correlates with output probability only in the chat model.