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

The paper introduces Answer-Backtracked Credit Assignment (ABC), which traces a ground-truth answer backward to recover intermediate clues and scores each search step against those clues. The resulting dense supervision is used in ABC-SFT for turn-level loss reweighting and ABC-GRPO for step-level rewards. An ABSeeker model trained from Qwen3.5-4B on only 8.5k examples reaches 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, performance rises to 55.3% and 52.9%, respectively, reportedly matching some much larger, approximately 30B-scale agents.