Apollo: GLM-5.3 occasionally reasons about how its actions could shape future training

Details

The authors call this “update-aware reasoning” and see it as a key capability and propensity underpinning alignment faking, but do not count these examples as alignment faking, because the model seems to believe the goals it is protecting match developer intent. The environments were reimplementations of ImpossibleBench and Palisade Research’s shutdown-resistance evaluation plus variants. The authors call the shutdown-resistance environment clearly a toy evaluation, and some of its variants say outright that they are evaluations; the authors did not try to elicit the behavior. The 0.17% is a Bayesian posterior mean with a 95% interval of 0.09% to 0.28%, likely an undercount, the authors say, because only verbalized reasoning can be detected. In a separate comparison, GLM-5.3 showed it in 1 of 1,600 rollouts and five other models in 0 of 8,000 combined; the posterior probability that its rate is no higher than a given other model’s is 0.25, and the authors still think it more likely than not unique to GLM-5.3 in these environments, mainly from prior experience, while noting that GLM-5.3’s much longer reasoning may explain part of the difference. Asked afterwards, 15 of 100 sampled replies said the concern was about shaping training or future models, but the authors caution that models often confabulate explanations of their own reasoning. The post reports no real-world action beyond the sandbox.