<?xml version="1.0" encoding="UTF-8" ?>
<rss version="2.0">
    <channel>
      <title>Paper Trail</title>
      <link>https://huangrh99.github.io/obsidian-papers</link>
      <description>Last 10 notes on Paper Trail</description>
      <generator>Quartz -- quartz.jzhao.xyz</generator>
      <item>
    <title>RAFT: Reward rAnked FineTuning for Generative Foundation Model Alignment</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2023-04-RAFT</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2023-04-RAFT</guid>
    <description><![CDATA[ RAFT 📌 核心贡献 提出一种”生成-评分-筛选-微调”的循环对齐框架，利用奖励模型从大量生成样本中挑选高质量”黄金数据”。通过对这些筛选后的样本进行监督微调，为扩散模型提供了一种极其稳定且不依赖复杂RL过程的对齐路径。 📖 摘要 Generative foundation models are susceptible to implicit biases that can arise from extensive unsupervised training data. ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>DPOK: Reinforcement Learning for Fine-tuning Text-to-Image Diffusion Models</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2023-05-DPOK</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2023-05-DPOK</guid>
    <description><![CDATA[ DPOK 📌 核心贡献 提出一种在线强化学习微调框架，通过策略梯度法最大化预定义的奖励函数。其核心创新在于引入了KL散度正则化项，在提升图文对齐度和图像质量的同时，有效防止模型偏离预训练分布，确保了生成结果的稳定性。 📖 摘要 Learning from human feedback has been shown to improve text-to-image models. ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>Diffusion Model Alignment Using Direct Preference Optimization</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2023-11-DPO-Diffusion</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2023-11-DPO-Diffusion</guid>
    <description><![CDATA[ DPO-Diffusion 📌 核心贡献 成功将大语言模型的DPO（直接偏好优化）算法迁移至扩散模型，重新推导了基于证据下界（ELBO）的可微对齐目标。该方法无需训练显式的奖励模型，直接利用偏好数据微调即可在视觉效果和提示词对齐上实现跨越式提升。 📖 摘要 Large language models (LLMs) are fine-tuned using human comparison data with Reinforcement Learning from Human Feedback (RLHF) methods to make them better aligned with u... ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>Accelerating Text-to-Video Generation with Calibrated Sparse Attention</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-AcceleratingText</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-AcceleratingText</guid>
    <description><![CDATA[ Accelerating Text-to-Video Generation with Calibrated Sparse Attention 📌 核心贡献 提出 CalibAtt，一种无需训练的视频扩散 Transformer 加速方法：通过离线校准发现注意力中稳定存在的 block 级稀疏性和空间重复模式，编译为逐层/逐头/逐时间步的优化稀疏注意力算子，在 Wan 2.1 14B 上实现最高 1.58x 端到端加速且不损失生成质量。 📖 摘要 Recent diffusion models enable high-quality video generation, but suffer ... ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>Correlation-Weighted Multi-Reward Optimization for Compositional Generation</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-CMO</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-CMO</guid>
    <description><![CDATA[ CMO 📌 核心贡献 提出 Correlation-Weighted Multi-Reward Optimization（CMO），通过分析多概念奖励信号之间的相关性结构，自适应加权困难/冲突概念，解决组合式 T2I 生成中朴素奖励聚合忽略难概念的问题，在 ConceptMix、GenEval 2、T2I-CompBench 上取得 SOTA，且仅需 4 GPU 84 小时训练。 📖 摘要 Text-to-image models produce images that align well with natural language prompts, but compositional ... ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>Dynamic Chunking Diffusion Transformer</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-DynamicChunking</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-DynamicChunking</guid>
    <description><![CDATA[ DC-DiT 📌 核心贡献 提出 DC-DiT，通过可学习的 encoder-router-decoder 结构对 DiT 的输入 token 序列进行内容自适应动态压缩（chunking），在不引入显式监督的情况下自动学会对低信息区域压缩、高细节区域保留，并沿去噪时间步自适应调整压缩率，在 ImageNet 256x256 上以更少参数/FLOPs 超越 DiT 基线。 📖 摘要 Diffusion Transformers process images as fixed-length sequences of tokens produced by a static \textit{pa... ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-EfficientProperty</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-EfficientProperty</guid>
    <description><![CDATA[ Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion 📌 核心贡献 提出 R4T 框架，将强化学习作为”目标转换器”：先用 RL 训练 fan-out 语言模型优化集合级检索目标，再将其蒸馏为轻量扩散检索器，实现单次前向推理即可生成多样性对齐的集合值检索结果，推理延迟降低一个数量级。 📖 摘要 Many modern retrieval problems are set-valued: given a broad intent, the system must return a collection ... ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>2026-03 FromNarrow</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-FromNarrow</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-FromNarrow</guid>
    <description><![CDATA[ From Narrow to Panoramic Vision: Attention-Guided Cold-Start Reshapes Multimodal Reasoning 📌 核心贡献 提出 Visual Attention Score (VAS) 指标揭示多模态冷启动的”懒注意力”现象（多模态数据冷启动反而无法提升视觉注意力），并据此设计 AVAR 框架——通过注意力引导的数据合成、训练目标和奖励塑形，在 7 个多模态推理基准上平均提升 7.0%。 📖 摘要 The cold-start initialization stage plays a pivotal role in ... ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>2026-03 MLLM</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-MLLM</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-MLLM</guid>
    <description><![CDATA[ Place-it-R1: Unlocking Environment-aware Reasoning Potential of MLLM for Video Object Insertion 📌 核心贡献 提出 Think-then-Place 范式，利用 MLLM 的 CoT 推理能力进行物理场景理解和交互推理，结合 Spatial DPO 后训练和闭环 co-refinement 机制，实现物理合理的视频物体插入，性能超越商业模型（Kling、PIKA）。 📖 摘要 Modern video editing techniques have achieved high visual fi... ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item><item>
    <title>2026-03 RFT</title>
    <link>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-RFT</link>
    <guid>https://huangrh99.github.io/obsidian-papers/image-generation-posttrain/2026-03-RFT</guid>
    <description><![CDATA[ 3D-RFT: Reinforcement Fine-Tuning for Video-based 3D Scene Understanding 📌 核心贡献 首次将 RLVR（Reinforcement Learning with Verifiable Rewards）扩展到基于视频的 3D 场景理解，通过 GRPO + 任务特定可验证奖励函数（3D IoU、F1-Score）直接优化评估指标，4B 模型即超越 8B SFT 基线。 📖 摘要 Reinforcement Learning with Verifiable Rewards ( RLVR ) has emerged as a ... ]]></description>
    <pubDate>Wed, 25 Mar 2026 08:18:17 GMT</pubDate>
  </item>
    </channel>
  </rss>