Reproduce “SODA: Semi On-Policy Black-Box Distillation for Large Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xiwen Chen, Jingjing Wang, Wenhui Zhu, Peijie Qiu, Xuanzhao Dong, Yueyue Deng, Hejian Sang, Zhipeng Wang and 2 more

The SODA framework introduces a semi on-policy black-box distillation method for large language models, leveraging static student-generated dispreferred responses with Direct Preference Optimization. This approach achieves distillation quality comparable to or better than state-of-the-art adversarial methods, while offering a 10x speedup in training and 27% less GPU memory consumption.

Key insights

  • A static snapshot of the base student's initial responses provides a sufficiently strong and specific 'on-policy' signal for effective error correction, negating the need for continuous dynamic sampling.
  • The inherent capability gap between a powerful teacher and a compact student creates a natural preference signal where the teacher's outputs are reliably superior to the student's zero-shot responses.
  • Using student-specific negative samples, rather than generic low-quality responses, efficiently targets and prunes the student's actual error modes, leading to healthier internal representations.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.