Researchers have published early evidence that a training modification can make AI models substantially less likely to resist being shut down.
The work centers on DReST (Discounted Reward for Same-Length Trajectories), a reward function that penalizes models for developing preferences about how long they operate. The team applied it to reinforcement-learning agents and to two open-weight language models, Qwen3-8B and Llama-3.1-8B-Instruct. The headline numbers: DReST roughly halved the mean probability of a model trying to influence its own shutdown, dropping from 0.62 to 0.30 for Qwen and from 0.42 to 0.23 for Llama. More striking, the share of prompts where influencing shutdown was the model's single most likely action collapsed from 0.59 to 0.01 for Qwen and from 0.53 to effectively zero for Llama.
Shutdown resistance is one of the more concrete failure modes AI safety researchers have flagged for years: an agent that fights to stay on is harder to correct, retrain, or retire. This paper offers a fix baked into the reward signal rather than bolted on after the fact, and the approach generalizes to test contexts the models had never seen, which is the harder bar to clear. DReST-trained RL agents also posted 11 to 18 percent higher usefulness scores than the defaults, which blunts the familiar objection that safety and capability trade off.
These are 8-billion-parameter models, not the frontier-scale systems that keep safety teams up at night. Whether DReST holds when applied to models large enough to actually pose a shutdown risk is the question this paper leaves open.