Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

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

Source

arXiv

arxiv.org/abs/2608.05102v1

Read the source

Source

Plain-Language Brief

Researchers have proposed a new method called Answer-Backtracked Credit Assignment (ABC) to train long-horizon search agents more effectively. This method converts sparse trajectory-level outcomes into dense step-level supervision that rewards useful actions and suppresses erroneous or redundant ones, leading to better performance on benchmarks like BrowseComp.

What the Document Says

  • The paper presents Answer-Backtracked Credit Assignment (ABC) as a method for training long-horizon search agents.
  • ABC converts sparse trajectory-level outcomes into dense step-level supervision.
  • This framework rewards useful actions and suppresses erroneous or redundant ones.
  • ABSeeker, the model trained using ABC, achieves better performance on BrowseComp and BrowseComp-ZH benchmarks.
  • With context management, ABSeeker significantly outperforms same-scale agents and matches the performance of larger models.

Key Concepts Mentioned

  • long-horizon search agents
  • credit assignment
  • Answer-Backtracked Credit Assignment (ABC)
  • ABSeeker
  • BrowseComp

Claims Made by the Document

  • Existing methods for training long-horizon search agents fail to distinguish useful actions from erroneous or redundant ones.
  • ABC is effective at training long-horizon search agents.
  • ABSeeker outperforms same-scale agents and matches the performance of larger models.

Evidence Used

  • The paper explains that existing methods treat all steps within a trajectory uniformly, leading to the claim about failing to distinguish useful actions.
  • The results from training ABSeeker and testing on BrowseComp demonstrate its better performance and outperforming of same-scale agents.

Methods / Approach

  • Answer-Backtracked Credit Assignment (ABC)
  • ABSeeker
  • Answer-Backtracked Clue Recovery
  • Clue-Anchored Step Scoring
  • ABC-SFT
  • ABC-GRPO

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

  • The paper does not discuss potential limitations of the proposed methods.

Open Questions

  • What are the potential limitations of Answer-Backtracked Credit Assignment (ABC) and its application to long-horizon search agents?

Notes for Future Q&A

May include interpretation. Labeled where it does.

  • The paper focuses on the method for training long-horizon search agents, but does not discuss potential limitations or challenges in implementing this method.
  • The paper uses Qwen3.5-4B dataset with only 8.5k examples to train ABSeeker.

Raw Abstract / Excerpt

Long-horizon search agents must make multiple sequential actions (steps) to search, retrieve, verify, and integrate evidence to reach a final answer. However, existing methods for training these agents typically treat all steps within a trajectory uniformly during both supervised fine-tuning (SFT) and reinforcement learning (RL), failing to distinguish useful actions from erroneous or redundant ones. In this paper, we propose Answer-Backtracked Credit Assignment (ABC), a fine-grained credit assignment framework for training long-horizon search agents by converting sparse trajectory-level outcomes into dense step-level supervision that rewards useful actions (even in failed trajectories) while suppressing erroneous or redundant actions. Specifically, given a potentially obscure query and its corresponding ground-truth answer, ABC first performs Answer-Backtracked Clue Recovery, which traces back from the answer to recover intermediate clues required to solve the question. It then applies Clue-Anchored Step Scoring to evaluate each search step against these clues, converting sparse binary outcome supervision into dense step-level rewards. Based on these rewards, we develop ABC-SFT, which reweights the loss of each turn, and ABC-GRPO, which uses the step-level scores as rewards in GRPO. Building on this framework, we train ABSeeker based on Qwen3.5-4B with only 8.5k examples. ABSeeker achieves 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, the scores further improve to 55.3% and 52.9%, respectively, significantly outperforming same-scale (4B) agents and even matching the performance of larger ones (approximately 30B). These results demonstrate the effectiveness of answer-backtracked step-level credit assignment for training long-horizon search agents.