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

TL;DR - ABSeeker is a long-horizon search agent trained with Answer-Backtracked Credit Assignment (ABC), which turns sparse trajectory-level outcomes into dense step-level supervision so useful search steps are rewarded and redundant or erroneous ones suppressed. It matters because a 4B model trained on just 8.5k examples reaches performance competitive with ~30B-scale agents on BrowseComp.

• ABC has two stages: Answer-Backtracked Clue Recovery traces from the ground-truth answer back to the intermediate clues needed to solve the query, then Clue-Anchored Step Scoring grades each search step against those clues to produce dense per-step rewards.
• The rewards feed two training variants: ABC-SFT reweights per-turn loss, and ABC-GRPO uses step-level scores as the RL reward signal; credit can be assigned to useful actions even inside failed trajectories.
• ABSeeker (Qwen3.5-4B, 8.5k training examples) scores 37.3% on BrowseComp and 39.1% on BrowseComp-ZH; adding context management raises these to 55.3% and 52.9%.
• Reported gains beat same-scale 4B agents and match roughly 30B-scale agents, arguing that fine-grained credit assignment substitutes for parameter count in long-horizon search.