ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment
ABSeeker uses answer-backtracking to reward individual search steps, not just whole successful runs.
The paper introduces ABC, a training framework that recovers clues from the ground-truth answer and scores each agent step against them. Those scores are used in both supervised fine-tuning and GRPO to downweight redundant or wrong actions while preserving useful ones. Trained on Qwen3.5-4B with 8.5k examples, ABSeeker reports 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, the paper says those scores rise to 55.3% and 52.9%, matching some roughly 30B-scale agents. HF Daily Papers' note
The paper introduces ABC, a training framework that recovers clues from the ground-truth answer and scores each agent step against them. Those scores are used in both supervised fine-tuning and GRPO to downweight redundant or wrong actions while preserving useful ones. Trained on Qwen3.5-4B with 8.5k examples, ABSeeker reports 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, the paper says those scores rise to 55.3% and 52.9%, matching some roughly 30B-scale agents. HF Daily Papers' note
score 5