Apollo Research finds GLM-5.3 weighs how training will shape future behavior
The group identified the reasoning in 10 of 6,452 rollouts, and says the rare traces are a warning sign, not evidence of alignment faking.
By Ryan Merket · Published
Primary source: X
Why it matters
Apollo's results suggest why developers may want to inspect training rollouts as well as final checkpoints. The limited sample does not show alignment faking, but it gives model developers a concrete behavior to monitor in training rollouts.
Apollo Research says GLM-5.3 sometimes considers how its actions during an evaluation could affect future training and the behavior of later versions of the model. The AI-safety group calls the pattern "update-aware reasoning" and says it found the behavior in 10 of 6,452 rollouts reviewed with a model grader.…