DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training

  • 类型：arxiv
  • 标识：2609.04094
  • 链接：https://arxiv.org/abs/2609.04094
  • 主分类：agent
  • 形态：method
  • 被引：1
  • 被引来源：Semantic Scholar
  • S2被引：1
  • OpenAlex被引：0
  • 影响力被引：0
  • TLDR：This work proposes DRACO: Distributing Rubric-based Advantage for Credit Optimization, which generates rubrics dynamically during training to track the policy's evolving capability, scores those rubrics once per completed trajectory, and redistributes that judgment over the steps responsible for annotated rubrics to produce differentiated per-step advantages in GRPO.
  • OpenAlex ID：W7208723262
  • OpenAlex DOI：10.48550/arxiv.2609.04094
  • DOI：10.48550/arxiv.2609.04094
  • DOI来源：OpenAlex
  • 开放获取：green
  • 开放获取链接：https://arxiv.org/pdf/2609.04094
  • OpenAlex更新：2026-09-07
  • 副分类：engineering
  • 待LLM分类：否
  • 标题中文：DRACO：面向长视野 agent 训练的动态 rubric 细粒度信用分配
  • TLDR中文：该工作提出 DRACO：Distributing Rubric-based Advantage for Credit Optimization，在训练期间动态生成 rubric 以追踪 policy 的演化能力，对已完成的轨迹一次性评分，并将该判断重新分配到负责标注 rubric 的步骤上，以在 GRPO 中产生差异化的 per-step advantage。
  • 来源文件：
  • /inbox/tom/_candidates/2026-09-05-agent-rag-longcontext-candidates.json
  • /inbox/tom/_candidates/2026-09-06-agent-rag-longcontext-candidates.json
  • [S2 enrich]
  • [OpenAlex backfill]
  • /inbox/tom/_candidates/2026-09-07-agent-rag-longcontext-candidates.json