---
title: "ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.05102v1"
permalink: "https://feed.astralor.com/items/itm_73b8a3bfe95a48d177274700"
published: "2026-08-05T17:41:31.000Z"
score: 76
channel: "ai"
tags: ["ABSeeker","arXiv","Qwen","Credit Assignment","Reinforcement Learning","Search Agent"]
---

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

## AI commentary

> What stands out here is moving search-agent training from whole-trajectory rewards to step-level credit assignment, especially identifying useful steps inside failed trajectories. For teams working on agent reasoning alignment or trying to make small models handle complex search tasks cheaply, this clue-backtracking approach is practical and worth adapting to their own pipelines.

## At a glance

Long-horizon search agents must perform multiple sequential actions to search, retrieve, verify, and integrate evidence. However, existing training methods treat all steps in a trajectory uniformly during both supervised fine-tuning (SFT) and reinforcement learning (RL), failing to distinguish useful actions from erroneous or redundant ones. This paper proposes Answer-Backtracked Credit Assignment (ABC), a fine-grained credit assignment framework that converts sparse trajectory-level outcomes into dense step-level supervision, rewarding useful actions even in failed trajectories while suppressing erroneous or redundant ones. ABC first performs Answer-Backtracked Clue Recovery, tracing 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 these clues. Based on these rewards, the authors develop ABC-SFT and ABC-GRPO, and train ABSeeker on Qwen3.5-4B with only 8.5k examples. ABSeeker achieves 37.3% on BrowseComp and 39.1% on BrowseComp-ZH. With context management, scores improve to 55.3% and 52.9%, significantly outperforming same-scale (4B) agents and even matching larger models around 30B. These results demonstrate the effectiveness of answer-backtracked step-level credit assignment for training long-horizon search agents.

## Key points

- The ABC framework converts sparse trajectory-level outcomes into dense step-level rewards, distinguishing useful, erroneous, and redundant actions.
- ABC-SFT reweights per-turn losses and ABC-GRPO uses step-level scores as rewards, both based on answer-backtracked clues.
- Trained on only 8.5k examples with a 4B model, ABSeeker reaches 37.3% on BrowseComp (55.3% with context management), approaching 30B-scale models.

## Full text

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.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_73b8a3bfe95a48d177274700