Long-horizon search agents face a fundamental challenge: they must execute multiple sequential steps—searching, retrieving, verifying, and integrating evidence—to reach a final answer. Yet existing training methods treat all steps uniformly, failing to distinguish useful actions from erroneous or redundant ones. This uniform treatment limits the efficiency and effectiveness of these agents.
In a new paper, researchers from Hugging Face propose Answer-Backtracked Credit Assignment (ABC), a fine-grained credit assignment framework that converts sparse trajectory-level outcomes into dense step-level supervision. ABC rewards useful actions even in failed trajectories while suppressing erroneous or redundant ones, providing a more nuanced training signal.
The framework operates in two stages: Answer-Backtracked Clue Recovery traces back from the ground-truth answer to recover intermediate clues needed to solve the question, and Clue-Anchored Step Scoring evaluates each search step against these clues, generating dense step-level rewards. Based on these rewards, the authors develop ABC-SFT (which reweights loss per turn) and ABC-GRPO (which uses step-level scores as rewards in GRPO).
Using this framework, the team trained ABSeeker on Qwen3.5-4B with only 8.5k examples. ABSeeker achieves 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, scores improve to 55.3% and 52.9%, respectively—significantly outperforming same-scale (4B) agents and even matching larger models of approximately 30B parameters.
The results demonstrate that answer-backtracked step-level credit assignment is a powerful approach for training long-horizon search agents, potentially enabling smaller models to compete with much larger counterparts. The code is available on GitHub, and the paper is on arXiv.