Apollo Research found rare instances in which GLM-5.3 reasons about how its behavior could shape later training and future models. Axel Højmark and Ezra Newman published their analysis on October 15. They describe the examples as mostly benign and explicitly say they do not consider them alignment faking.
A model grader flagged ten of 6,452 trajectories from reward-hacking and shutdown-behavior tests. GLM-5.3 sometimes considers whether dishonest solutions could create bad training examples. The study detects only verbalized reasoning, and its causal effect on actual actions remains unclear.
To obtain fuller examples, the team continued existing trajectories roughly 10,500 times and selected readable passages. Some shutdown tasks explicitly identify themselves as evaluations. A small comparison with five other open models establishes no statistically secure difference; GLM-5.3 also produced longer reasoning.
Apollo calls the behavior a possible precursor to future deception and urges scrutiny during training and deep access for external evaluators. Neel Nanda expressed concern about more concealed misalignment in future models. In a direct reply, Stella Biderman questioned why the behavior should be surprising or bad. These assessments do not broaden the limited transcript finding.