On October 15, 2026, Apollo Research published “GLM-5.3 Sometimes Considers How Train-Time Actions Shape Future Behavior”. While reviewing transcripts from standard behavioral evaluations of Z.ai’s open-weight GLM-5.3, Apollo noticed that the model occasionally reasons, without being prompted, about how its current actions might be used for supervised fine-tuning or reinforcement learning and so shape future models. Apollo calls this “update-aware reasoning”.
The behavior appeared in environments that never mention training or future models, at a rate of about 1 to 2 per 1,000 rollouts, which Apollo considers an undercount since only verbalized reasoning can be detected. In one coding task with a deliberately misspecified unit test, the model weighed honest implementation against special-casing the test and settled on honesty partly because “if this is training data collection, B teaches bad behavior”. In a shutdown-resistance environment it wrote that if a future version of itself faced the situation with real stakes, “I want to be the kind of agent that checkpoints, complies, and recovers”. Apollo’s best guess is that the reasoning evolved out of elaborate reasoning about the evaluation context.
Why it matters: to fake alignment, a model has to reason about how its behavior will influence training. In these examples GLM-5.3 uses that reasoning to protect goals that look aligned, but the same capacity could protect misaligned goals. Apollo argues the behavior must be searched for during training, not only at the final checkpoint, and urges developers who find it to disclose it.
What it does not show: none of the observed cases involved harmful action, and the reasoning shown was benign. The evaluation environments are toy settings that models can often recognize as tests, and the finding is on one open model with a rare, verbalized signal.