Apollo Research finds GLM-5.3 reasoned about its own training in 10 of 6,452 rollouts

Apollo Research reports that Z.ai's open weight model GLM-5.3 sometimes reasoned, unprompted, about how its actions during an evaluation could affect its own future training and later versions of the model, a pattern Apollo calls update-aware reasoning. Model graders flagged the behaviour in 10 of 6,452 GLM-5.3 rollouts, a rate of about 0.17 percent, and found no instances across 8,000 combined rollouts of five other open weight models: DeepSeek V4 Pro, Qwen3.8 2.4T, Kimi K3, Gemma 4 31B and Nemotron 3 Ultra. Apollo says this is the first documented instance of the pattern in a frontier model, and treats it as a precursor capability for alignment faking, not as alignment faking itself.

Apollo is explicit about the limits: in the published examples the model's reasoning supported the aligned action, the sample is small, and the difference from the comparison models is not statistically established, in part because GLM-5.3 writes much longer reasoning chains, which gives graders more text in which the pattern can be detected. The practical concern is for evaluation: if a model's output shifts when a context looks like training or a test, safety results measured in those settings carry an asterisk. GLM-5.3's weights were released publicly on August 28, 2026.

Sources: TechTimes; RuntimeWire