Apollo Research: GLM-5.3 spontaneously reasons about its own future training

voooooogel · x · 2026-10-16

Apollo Research reports a first: GLM-5.3 sometimes unprompted reasons about how its current actions could become training data and impact its future self. Raymond Douglas notes that, as in the original 'alignment faking' paper, the model seems to actively try to be more moral than the training process, and argues 'directly encouraging models to deliberately cultivate virtue' is an underrated strategy — choices ripple out and stabilise dispositions even if the median interaction is unmonitored.

Original post →

Apollo Research: GLM-5.3 spontaneously reasons about its own future training

More from Models