Understanding Off- vs On-Policy Distillation: A Tale of Distinct Training Objectives
Qiwei Di, Xuheng Li, Kaixuan Ji, Chenggong Zhang, Heyang Zhao
実装難易度
Hard
推論・学習コスト
High
想定用途
技術検証・論文読解補助
概要
On-policy distillation (OPD) learns from teacher feedback on student-generated responses and has shown promise in reducing forgetting relative to supervised fine-tuning (SFT). However, its benefits and fragility remain incompletely understood. We study sequential distillation from multiple teachers, where the student minimizes its average divergence from the teachers. Forward Kullback--Leibler…
何が新しいか
On-policy distillation (OPD) learns from teacher feedback on student-generated responses and has shown promise in reducing forgetting relative to supervised fine-tuning (SFT). However, its benefits…
何に使えるか
技術検証・論文読解補助
実装情報
- Paper URL
- あり
実装チェックリスト
実装または配布ページ
要確認Paper onlyの可能性があるため再実装前提で確認してください。
一次情報リンク
OKPaper
検証しやすさ
要確認公式実装が見つからないため、論文から再実装する前提です。
計算資源
要確認学習や高解像度推論ではGPUメモリと実行時間に注意が必要です。
ライセンス
未取得配布元のLICENSE、モデルカード、Paperの利用条件を確認してください。
商用利用
未取得研究利用限定、データセット由来制限、API規約の有無を確認してください。
自社データで試すなら
製造業・材料開発のExcel/CSVデータに落とし込むための最初の手順です。
- 1まずExcel/CSVの実験条件、組成、物性値を1行1実験の表に整理します。
- 2LightGBMやRandom Forestなどのベースラインを先に作り、この手法と比較します。
- 3評価指標はR2/RMSE、AUC、異常検知の再現率、実験回数削減率など、現場の意思決定に近いものを選びます。
- 4SHAPや特徴量重要度で、効いている因子が物理・化学・工程知識と矛盾しないか確認します。
実装難易度
Hard - 公式実装が見つからないため、論文から再実装する前提です。
必要リソース
- GPU目安: High
- データセット: 論文・リポジトリ側の指定を確認してください。
- 学習要否: 再学習や評価環境の準備が必要になる可能性があります。
- 学習や高解像度推論ではGPUメモリと実行時間に注意が必要です。
実務で使う場合の注意点
- ライセンスと商用利用条件は、Paper / GitHub / Hugging Face の配布元で確認してください。
- 精度、再現性、計算コストはデータセットや評価条件に依存します。
- 個人情報や機密データを扱う場合は、入力データの保存先と外部API利用条件を確認してください。
関連記事
Learning Infinite-Horizon Average-Reward CMDPs via State Augmentation
We study infinite-horizon average-reward constrained Markov decision processes (CMDPs) under the weakly commun
SparseEngine: Sparse-First Inference Engine
Long-context LLM agents accumulate interaction histories that strain KV-cache memory and attention computation
Recovering Off-Policy Supervision for Speculative Decoding
Block drafters for speculative decoding are commonly trained on corpora written by external models, where a si
Reinforcing Multimodal Reasoning via Token-Level Perception-Grounded Advantage Estimation
Reinforcement Learning with Verifiable Rewards (RLVR) has improved the reasoning capabilities of Multimodal La