代理訓練與深度搜尋
ABSeeker Backtracks from Correct Answers to Derive Search Clues, Raising a 4B Model’s BrowseComp Score to 55.3%
ABSeeker breaks down a search agent’s final success or failure into step-level rewards, preserving useful queries from failed trajectories while downweighting erroneous steps in successful ones. The team has released the 4B model weights and the inference and training pipelines, though its best results still rely on additional context management and an LLM judge.
Long-horizon search agents typically receive only a binary “correct or incorrect” reward at the end. As a result, one wrong answer can erase credit for useful evidence found along the way, while one correct answer may reinforce redundant queries and faulty reasoning. ABSeeker’s Answer-Backtracked Credit Assignment (ABC) first uses the known correct answer to search backward, reconstructing the entities, facts, and relationships that should have been discovered. Clue-Anchored Step Scoring then checks whether each turn finds or verifies a clue, eliminates an incorrect candidate, or introduces an error.
These scores feed into two training stages: ABC-SFT reweights the supervised loss according to per-turn scores, while ABC-GRPO converts step-level scores into reinforcement learning rewards. Using Qwen3.5-4B as the base model, the team performed SFT on 8,500 OpenSeeker trajectories, including 3,000 failed runs. RL then used 1,000 questions with eight rollouts per question and a maximum of 200 search turns per rollout. The released code includes pipelines for ReAct inference, clue recovery, step-level scoring, Slime SFT, and veRL GRPO. Search and webpage retrieval rely on Serper and Jina, respectively, along with an additional summarization model.
Without context management, ABSeeker scored 37.3%, 39.1%, 77.0%, 46.0%, and 81.6% on BrowseComp, BrowseComp-ZH, xbench-2505, xbench-2510, and GAIA-text, respectively. With context management applied, its BrowseComp and BrowseComp-ZH scores rose to 55.3% and 52.9%. In ablation experiments, ABC-GRPO outperformed standard GRPO trained on the same data across all five benchmarks. However, the clue and step scores are still generated by an LLM, potentially encoding answer leakage, judge bias, or single-path assumptions into the reward. The validation set was also sampled, and final answers were judged by DeepSeek-V4-Flash. Future areas to watch include reproducibility across search engines, real-world API costs, and whether the method can still benefit research tasks that lack a single answer that can be traced backward.