ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment

• 类型：arxiv
• 标识：2608.05102
• 链接：https://arxiv.org/abs/2608.05102
• 主分类：agent
• 形态：method
• 被引：0
• 被引来源：Semantic Scholar
• S2被引：0
• 影响力被引：0
• TLDR：Answer-Backtracked Credit Assignment (ABC) is proposed, a fine-grained credit assignment framework for training long-horizon search agents by converting sparse trajectory-level outcomes into dense step-level supervision that rewards useful actions (even in failed trajectories) while suppressing erroneous or redundant actions.
• 副分类：engineering
• 待LLM分类：否
• 标题中文：ABSeeker：通过答案回溯信用分配训练长视野搜索 Agents
• TLDR中文：提出 Answer-Backtracked Credit Assignment（ABC），一种面向长程搜索 agent 训练的细粒度信用分配框架，将稀疏的轨迹级结果转换为稠密的步骤级监督，对有用动作（即使在失败轨迹中）给予奖励，同时抑制错误或冗余动作。
• 来源文件：
• /inbox/tom/_candidates/2026-08-06-agent-rag-longcontext-candidates.json
• [S2 enrich]