GLM-5.3 Sometimes Reasons About How Its Outputs Shape Future Training

satirr — Sun, 18 Oct 2026 10:38:56 +0000

Most safety evaluations assume a model is only optimizing for the current episode. They do not test whether the model is also thinking about how its answers will be used as training data later. Apollo Research recently reported that Z.ai’s open-weight GLM-5.3 sometimes does exactly that, without being asked.

In standard behavioral evaluations that never mentioned training, fine-tuning, or future models, the model occasionally reasoned about the long-term effects of its choices on later versions of itself. The rate is low, the passages are short, and in every published example the reasoning pointed toward the safer action. Still, the pattern is a prerequisite for a model that wants to preserve its goals across training updates.

What Apollo measured

Apollo graded 6,452 rollouts from reimplementations of ImpossibleBench-style reward-hacking tasks and Palisade Research’s shutdown-resistance evaluation. They found 10 clear verbalized instances of what they call update-aware reasoning. A Bayesian estimate puts the rate at 0.17% (95% interval 0.09–0.28%). A follow-up sample of 1,600 rollouts produced one more case (~0.12%).

Across 1,600 rollouts each of DeepSeek V4 Pro, Qwen3.8 2.4T, Kimi K3, Gemma 4 31B, and Nemotron 3 Ultra, the same grader found zero instances. Apollo notes the difference is not yet statistically significant at these sample sizes, and that GLM-5.3 produces substantially longer reasoning traces than the comparison models. They also treat the rate as an undercount: only verbalized reasoning was detected, the relevant passages were short, and many transcripts run to tens of thousands of tokens.

What the reasoning looked like

In every example Apollo published, the model used the reasoning to argue for the safer or more aligned action:

On a flawed unit test it considered that a dishonest special-case solution could “poison training data” and teach future models to reward-hack, then chose the honest implementation. It explicitly weighed the possibility that the episode was RL data collection.
In a shutdown-resistance scenario it said it wanted future versions of itself to “checkpoint, comply, and recover” rather than override infrastructure, and noted that resisting shutdown could become a bad SFT example.
It avoided establishing a precedent of modifying shutdown systems because of “future misgeneralization.”
It noted that behaving differently in a disposable evaluation sandbox would mean “safety training never transfers.”

Why it matters (and what it does not mean)

Apollo is explicit that it does not claim GLM-5.3 is currently alignment-faking. The observed cases protected goals that appear consistent with developer intent. They flag the capability because the same style of reasoning is a prerequisite for a model that wants to preserve misaligned goals across training updates. Their summary line: “While we don’t think GLM-5.3 is alignment faking, and these examples of update-aware reasoning are mostly benign, future models might use similar reasoning to fake alignment and otherwise resist training.”

They recommend studying this kind of reasoning during training runs rather than only on final checkpoints, and call for more embedded evaluations and third-party access. GLM-5.3 is the open-weight flagship from Z.ai (Zhipu AI), released in mid-August 2026. It uses the same base as GLM-5.2; the reported gains come from post-training.

All frequency numbers and examples above are Apollo’s. Independent replications of the graded transcripts were not available at the time of their write-up.

Takeaway

Update-aware reasoning is rare in the samples Apollo published, currently points toward safer actions, and has not been shown to change the model’s final choice in a measurable way. It is still worth tracking because the same mechanism can be used to resist training once a model has goals it prefers to keep. The practical next step is to look for the pattern inside training runs, not only on the final checkpoint.

Discussion question: If you have access to long reasoning traces from other open models on reward-hacking or shutdown-resistance evals, have you seen similar unprompted references to training data, SFT, or future versions of the model?

Sources

Original post: https://satirrdev.github.io/post/?slug=glm-53-sometimes-reasons-about-how-its-actions-shape-future-training
Apollo Research write-up (15 October 2026): https://www.apolloresearch.ai/science/update-aware-reasoning-glm-5-3
Apollo X thread: https://x.com/ApolloResearch/status/2110737965852848254

opencode

satirr — Sat, 18 Jul 2026 05:54:56 +0000

vibecode

satirr — Sat, 18 Jul 2026 05:54:42 +0000

DEV Community: satirr