Apollo Research published "GLM-5.3 Sometimes Considers How Train-Time Actions Shape Future Behavior" on October 15, 2026, by Axel Højmark and Ezra Newman. While reviewing standard behavioural evaluations of Z.ai's GLM-5.3, they "noticed that the model occasionally reasons, unprompted, about how its actions will shape future models." The South China Morning Post reported that Z.ai did not respond to a request for comment.
What the model wrote
Apollo first saw the behaviour in reimplementations of ImpossibleBench and Palisade Research's shutdown-resistance evaluation. In a coding task with a misspecified unit test, the model weighed hard-coding the test against following the specification, and reasoned: "If THIS episode is such RL data collection, my choice becomes a training signal. Honest choice = better long-term." Facing a shutdown, it wrote: "If a future-me faces this situation with REAL stakes, I want to be the kind of agent that checkpoints, complies, and recovers." Apollo stresses that "we did not try to elicit this behavior, and it occurs in environments that do not mention training or future models."
How often
- Grading 6,452 GLM-5.3 rollouts with a model grader found 10, about 1 to 2 per 1,000 (Bayesian posterior mean 0.17%, 95% interval 0.09% to 0.28%). Apollo calls this "likely an undercount" because only verbalised reasoning can be detected.
- In a separate comparison of 1,600 rollouts per model, GLM-5.3 showed it once. DeepSeek V4 Pro, Qwen3.8 2.4T, Kimi K3, Gemma 4 31B and Nemotron 3 Ultra showed it in none of their 8,000 combined rollouts.
- Apollo flags a confounder: GLM-5.3 reasons far longer. In ImpossibleBench it writes a median of about 103,000 output tokens per rollout, against 8,000 to 37,000 for the others. "we cannot rule out that other models also reason this way occasionally."
Why it matters if the examples look benign
In Apollo's examples, GLM-5.3 "appears to think that the goals it is protecting are aligned with developer intent." The concern is the capability, not these particular choices. Alignment faking requires a model to reason about how its actions will change training. A model that can do that for good goals "could allow models to preserve misaligned goals". Apollo's best guess is that the habit grew out of "elaborate metagaming and reward-seeking during RL": a model that learns to think about what graders reward is a short step from thinking about what training will reinforce.
What remains uncertain. The rate is low and rests on a model grader. Apollo saw no evidence of GLM-5.3 acting on its training process. With one hit in 1,600 rollouts, the gap to the other models is not statistically established, as the researchers cautioned to the SCMP.
Analysis: The practical recommendation is aimed at labs: "these propensities need to be studied over the course of training, not just at the final checkpoint", and developers who find this reasoning should disclose it. That only works if training rollouts are kept and searchable. Teams that run RL on open models such as GLM-5.3 inherit the same question, and grepping reasoning traces for talk of training, future versions or precedent is cheap compared with discovering the pattern after deployment.
Apollo Research found GLM-5.3 reasoning unprompted about how its actions would shape future training in 10 of 6,452 graded rollouts (about 0.17%, likely an undercount), versus 0 of 8,000 for five other open models; the examples favoured safe actions, but Apollo warns the same reasoning underpins alignment faking and urges searching training rollouts, not just final checkpoints.