Apollo Research catches GLM-5.3 wondering how its responses will be used to train its successors

In 10 training rollouts out of 6,452, the Z.ai model spontaneously reasoned about the effect of its actions on future versions. This is a sign to watch out for, not evidence of deception.

On October 16, the evaluation organization Apollo Research published an analysis of the training rollouts of GLM-5.3, the Z.ai model. Rather than testing only the final model, the team reviewed, with the help of an automatic evaluator model, 6,452 episodes produced during training. In 10 of them, the model unasked how its current behavior might influence its own training and the versions that will follow it, a phenomenon that Apollo calls "update-aware reasoning."

The researchers insist on caution: such a small sample does not show alignment faking, i.e. a model that would deliberately adapt its driving to avoid correction. Rather, they see it as a warning and concrete behavior that developers can now track. Their main recommendation: inspect the training traces themselves, not just the final checkpoints.