ABSeeker Improves Long-Horizon Search Agents with Backtracked Credit

Training long-horizon search agents, which require multiple sequential actions to gather, verify, and integrate evidence, is challenging because traditional methods treat all steps in a trajectory uniformly. This fails to differentiate between useful, erroneous, or redundant actions, even in trajectories that ultimately fail. A new framework, Answer-Backtracked Credit Assignment (ABC), addresses this by providing fine-grained, step-level supervision. Given a query and its ground-truth answer, ABC first performs "Answer-Backtracked Clue Recovery" to identify intermediate clues needed to solve the question. It then applies "Clue-Anchored Step Scoring" to evaluate each search step against these clues, transforming sparse binary outcomes into dense, step-level rewards. Building on ABC, the ABSeeker agent, trained on Qwen3.5-4B with a relatively small dataset, achieved significant performance improvements. It outperformed same-scale agents and even matched larger 30B models on benchmarks like BrowseComp, demonstrating the effectiveness of this step-level credit assignment for training highly capable long-horizon search agents.