<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>GFlowNet on MessageDaily</title><link>https://inkeast.github.io/MessageDaily/tags/gflownet/</link><description>Recent content in GFlowNet on MessageDaily</description><generator>Hugo</generator><language>zh-cn</language><lastBuildDate>Wed, 09 Sep 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://inkeast.github.io/MessageDaily/tags/gflownet/index.xml" rel="self" type="application/rss+xml"/><item><title>FlowBalance：验证器锚定的自改进——把符号门控装进轨迹平衡 精读</title><link>https://inkeast.github.io/MessageDaily/posts/2026-09-09-flowbalance-verifier-grounded-self-improvement-paper-reading/</link><pubDate>Wed, 09 Sep 2026 00:00:00 +0000</pubDate><guid>https://inkeast.github.io/MessageDaily/posts/2026-09-09-flowbalance-verifier-grounded-self-improvement-paper-reading/</guid><description>腾讯 HY LLM Frontier×UPenn 推出 FlowBalance：冻结的特权后见视角对已采样轨迹打密集分，验证器组相对优势用符号门控决定引导方向（正保留/负反转/零关闭），profiled trajectory balance 拟合归一化目标分布。Qwen3-8B 五基准平均 67.61 超 GRPO/OPSD/RLSD/FlowRL，AIME24 达标 100 步 vs GRPO 143 步，四条目标级定理精确刻画反自确认机制。</description></item></channel></rss>