On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity
JA title: サンプル実演を用いた方針内自己蒸留は出力多様性を低下させる
What to read for
3 checks before implementation
- 01Article focus
AI が学習する際に「正しい答え例を見せながら自分で教え込む」という手法を使うと、精度は上がりますが、異なるパターンの答えが減ってしまう問題があります。
- 02Operational lens
理論的研究であり、実装・導入は直接的ではありません。
- 03Verify in source
arXiv (cs.AI) (Research): check the dataset, evaluation setup, and adoption assumptions for Code Generation.
On-policy self-distillation achieves strong pass@1 accuracy by using a single model as both teacher and student, with the teacher conditioned on a correct demonstration to provide dense token-level feedback. We show that this could come at a hidden cost: rollout diversity decreases and pass@k curves flatten (i.e., generating more rollouts fails to improve accuracy). We trace this to compounding biases in the design of self-distillation with sampled demonstrations. The teache…
ゼロビズAX View — How Japanese companies can apply it
理論的研究であり、実装・導入は直接的ではありません。ただし LLM ファインチューニング環境で推論多様性が重要な場合（レコメンデーション・創作支援など）は参考になります。既存の自己蒸留パイプラインを検証する際に「出力多様性の低下」を指標として監視することをお勧めします。情報不足のため具体的な実装方法・費用感は判断不能。
Plain-language glossary
Simple explanations of the technical terms used in this article.
- トークン
- The small chunks an AI breaks text into. Usage is often billed by the number of tokens.
- LLM（大規模言語モデル）
- The 'brain' behind chat AIs — a model trained on huge amounts of text to understand and write language.
- ファインチューニング（追加学習）
- Training an existing AI further on your own data so it fits your specific use case.
- 推論（インファレンス）
- The step where a trained AI actually answers or predicts — running the model in production.
Next step
Apply this article to your context
RAG, AI agents, and generative AI APIs implemented around your operations.
Business AI
Primary source: https://arxiv.org/abs/2606.26091v1
This article is an AI-generated summary of overseas primary sources. Translations or summaries may be imperfect — always confirm the original before making implementation decisions. The "ゼロビズAX View" is our applied interpretation and does not guarantee any specific outcome.
More in Global AI →← Back to list