ABSeeker: New Framework for Training Long-Horizon Search Agents

A recent study published on arXiv (2608.05102) presents the Answer-Backtracked Credit Assignment (ABC) framework designed for training agents that engage in long-horizon searches. These agents execute a series of actions to search, retrieve, verify, and synthesize evidence. Current approaches do not differentiate between beneficial actions and those that are incorrect or unnecessary during supervised fine-tuning and reinforcement learning. ABC transforms sparse outcomes at the trajectory level into detailed supervision at the step level, incentivizing valuable actions even in unsuccessful attempts while minimizing the impact of mistakes. Additionally, the framework features Answer-Backtracked Clue Recovery, which retraces steps from the answer to uncover intermediate clues. The authors of the paper are researchers who aim to enhance the effectiveness and precision of search agents in intricate tasks.