Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation
Summary
论文提出 D-OPCD（Diffusion On-Policy Context Distillation），将文本生成图像智能体框架（harness）通过记忆、技能和迭代优化获得的增益蒸馏进扩散模型权重中，使模型在仅使用原始查询时也能保留部分收益，平均直接生成分数从 60.52 提升至 65.09，并通过 AutoSkillEvolver 实现框架与模型的持续协同进化。
View Cached Full Text
Cached at: 10/08/26, 12:14 PM
Paper page - Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation
Source: https://huggingface.co/papers/2610.07250
Abstract
WrappinganimagegenerationmodelinanagenticharnesscaneffectivelyboostText-to-Imagetaskperformance:theharnesscanleveragememory,skills,workfloworchestration,resultverification,anditerativerefinementtocontinuallyconstructandreviseprompts,therebyelicitingbetterimages.Thesegains,however,remainexternaltothediffusionmodelandarerealizedonlywhilethefullharnessruns.WeproposeDiffusionOn-PolicyContextDistillation(D-OPCD),whichtreatstheagent-improvedpromptasprivilegedcontextanddistillstheknowledgeencodedintheagentharnessintotheweightsofthediffusionmodel,sothatthemodelretainspartoftheharness’sbenefitwhenconditionedontheoriginalqueryalone.UsingaText-to-ImageagentequippedwithourproposedAutoSkillEvolver(ASE),weshowthatD-OPCDcaninternalizeharnesscapabilitiesintothegenerator’sweights,raisingtheaveragedirect-generationscorefrom60.52to65.09acrossfourbenchmarks.Withthisknowledgeabsorbedintotheweights,theharnesscansheditssaturatedskillsandresumeevolving:asecondASEroundontheupdatedgeneratorimprovesonaskill-freeharnessbyadditional1.83points,pointingtowardtext-to-imagesystemsinwhichharnessandmodelkeepimprovingeachotherthroughcontinualco-evolution.
View arXiv pageView PDFProject pageGitHub0Add to collection
Get this paper in your agent:
hf papers read 2610\.07250
Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash
Models citing this paper0
No model linking this paper
Cite arxiv.org/abs/2610.07250 in a model README.md to link it from this page.
Datasets citing this paper0
No dataset linking this paper
Cite arxiv.org/abs/2610.07250 in a dataset README.md to link it from this page.
Spaces citing this paper0
No Space linking this paper
Cite arxiv.org/abs/2610.07250 in a Space README.md to link it from this page.
Collections including this paper0
No Collection including this paper
Add this paper to acollectionto link it from this page.
Similar Articles
Internalizing Agent Experience into Diffusion Model Weights via On-Policy Context Distillation
该论文提出 D-OPCD（Diffusion On-Policy Context Distillation），将 T2I 智能体 harness 通过提示词积累的经验蒸馏进扩散模型权重中，使模型在仅用原始查询条件时也能保留部分 harness 增益，并结合 Auto Skill Evolver 实现 harness 与模型的持续共同进化。
DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models
DiffusionOPD proposes a multi-task training paradigm for diffusion models that uses online policy distillation to efficiently combine task-specific teachers into a unified student, achieving state-of-the-art results on all evaluated benchmarks.
Simplifying and Improving On-Policy Distillation for Few-Step Diffusion
提出简化匹配蒸馏（SMD）框架，通过纯回归的on-policy速度回归与Pivot残差代理，实现无需辅助评分模型或对抗训练的4步少步扩散生成，在图像、3D、视频及音视频联合生成任务上超越强基线并可扩展至33B参数。
Data-Efficient Autoregressive-to-Diffusion Language Models via On-Policy Distillation
The paper introduces OPDLM, a method that transforms autoregressive language models into diffusion language models via on-policy distillation, requiring 15x to 7000x fewer training tokens while retaining knowledge from the original model.
Correction Training for Diffusion Models
论文提出一种针对扩散模型的“修正训练”框架，通过训练一个以基模型自身预测为输入的修正器来缩小训练-推理差距，显著提升无引导/无条件生成的样本质量，并可通过蒸馏将收益无损保留到基模型中。