Researchers Yijun Lu, Rui Ye, Jiajun Wang, Yuwen Du, Tian Jin, Songhua Liu and Siheng Chen released ABSeeker, a long-horizon search agent trained with Answer-Backtracked Credit Assignment, on Hugging Face Papers after publication on Aug 5 and submission by RuiYe on Aug 6. The paper says long-horizon search agents must take multiple sequential actions to search, retrieve, verify and integrate evidence before producing a final answer, but existing supervised fine-tuning and reinforcement learning methods usually treat every step in a trajectory the same way.

Answer-Backtracked Credit Assignment, or ABC, is designed to assign credit at the step level instead of judging only the whole search path. Given a potentially obscure query and its ground-truth answer, ABC first performs Answer-Backtracked Clue Recovery, which traces backward from the answer to recover intermediate clues needed to solve the question. It then applies Clue-Anchored Step Scoring to evaluate each search step against those clues, turning sparse binary outcome supervision into dense step-level rewards.

The authors use those rewards in two training variants. ABC-SFT reweights the loss of each turn during supervised fine-tuning, while ABC-GRPO uses the step-level scores as rewards in GRPO. The approach rewards useful actions, including useful actions inside failed trajectories, and suppresses erroneous or redundant actions that may appear during a multi-step search process.

ABSeeker was trained on Qwen3.5-4B with only 8.5k examples. The system achieved 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, the scores rose to 55.3% and 52.9%, respectively, which the paper says significantly outperformed same-scale 4B agents and matched the performance of larger agents at approximately 30B.

Researchers Yijun Lu, Rui Ye, Jiajun Wang, Yuwen Du, Tian Jin, Songhua Liu and Siheng Chen released ABSeeker, a long-horizon search agent trained with Answer-Backtracked Credit Assignment, on Hugging Face Papers after publication on Aug 5 and submission by RuiYe on Aug 6. The paper says long-horizon search agents must take multiple sequential actions to search, retrieve, verify and integrate evidence before producing a final answer, but existing supervised fine-tuning and reinforcement learning methods usually treat every step in a trajectory the same way.

Answer-Backtracked Credit Assignment, or ABC, is designed to assign credit at the step level instead of judging only the whole search path. Given a potentially obscure query and its ground-truth answer, ABC first performs Answer-Backtracked Clue Recovery, which traces backward from the answer to recover intermediate clues needed to solve the question. It then applies Clue-Anchored Step Scoring to evaluate each search step against those clues, turning sparse binary outcome supervision into dense step-level rewards.

The authors use those rewards in two training variants. ABC-SFT reweights the loss of each turn during supervised fine-tuning, while ABC-GRPO uses the step-level scores as rewards in GRPO. The approach rewards useful actions, including useful actions inside failed trajectories, and suppresses erroneous or redundant actions that may appear during a multi-step search process.

ABSeeker was trained on Qwen3.5-4B with only 8.5k examples. The system achieved 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, the scores rose to 55.3% and 52.9%, respectively, which the paper says significantly outperformed same-scale 4B agents and matched the performance of larger agents at approximately 30B.