The Year in AI Papers
The year/Independent research

Paper 2511.10643

Black-Box On-Policy Distillation of Large Language Models

Tianzhu Ye, Li Dong, Zewen Chi, Xun Wu, Shaohan Huang, Furu Wei

Published
Nov 2025
Research lab
Independent
Citations
43
GitHub
Not linked

01 In brief

Summary

The paper introduces Generative Adversarial Distillation (GAD), a method for black-box distillation of large language models (LLMs) where only the teacher's text outputs are accessible.

GAD frames the student as a generator and trains a discriminator to distinguish student from teacher responses, forming a minimax game.

The discriminator acts as an on-policy reward model that co-evolves with the student, providing stable feedback.

Experiments using GPT-5-Chat as teacher and Qwen2.5 and Llama3 students show GAD consistently outperforms sequence-level knowledge distillation (SeqKD) on LMSYS-Chat and out-of-distribution benchmarks.

Notably, Qwen2.5-14B-Instruct trained with GAD approaches the teacher's performance.

Ablations show the importance of warmup, Bradley-Terry loss, and equal generator-discriminator sizes.

GAD avoids reward hacking and overfitting to local patterns, demonstrating superior generalization.

02 From the paper

Abstract

Black-box distillation creates student large language models (LLMs) by learning from a proprietary teacher model's text outputs alone, without access to its internal logits or parameters. In this work, we introduce Generative Adversarial Distillation (GAD), which enables on-policy and black-box distillation. GAD frames the student LLM as a generator and trains a discriminator to distinguish its responses from the teacher LLM's, creating a minimax game. The discriminator acts as an on-policy reward model that co-evolves with the student, providing stable, adaptive feedback. Experimental results show that GAD consistently surpasses the commonly used sequence-level knowledge distillation. In particular, Qwen2.5-14B-Instruct (student) trained with GAD becomes comparable to its teacher, GPT-5-Chat, on the LMSYS-Chat automatic evaluation. The results establish GAD as a promising and effective paradigm for black-box LLM distillation.