ABSeeker is a long-horizon search agent trained with Answer-Backtracked Credit Assignment (ABC), a fine-grained credit assignment framework that converts sparse trajectory-level outcomes into dense step-level supervision. We trained ABSeeker based on Qwen3.5-4B with only 8.5K training examples and achieved strong performance on long-horizon search benchmarks:
teasor
Highlights
Fine-grained credit assignment: ABC rewards useful actions, even in failed trajectories, while suppressing erroneous or redundant actions.
Unified step-level supervision: ABC-SFT reweights the loss of each turn, while ABC-GRPO uses step-level scores as rewards during reinforcement learning.
Strong performance with only 4B parameters: ABSeeker achieves 37.3% on BrowseComp and 39.1% on BrowseComp-ZH, further improving to 55.3% and 52.9% with context management and remaining competitive with substantially larger search agents.
