GAD proposes on-policy distillation for black-box LLMs
An article dated September 13, 2026 presents a method that learns from a proprietary teacher model’s text outputs, without access to logits or parameters.
Published on September 13, 2026, the article presents Generative Adversarial Distillation (GAD), an on-policy, black-box distillation method for language models. The proposal uses only text outputs from a proprietary teacher model.
In the described method, a discriminator distinguishes student responses from teacher responses in a minimax game. This enables distillation without access to the teacher’s logits or parameters. To assess the proposal’s scope and evidence, consult the original article by title and check its method and reported results; the available summary does not specify quantitative results.