ABSeeker: A 4B Search Agent That Fights Like a 30B
Top agent paper on Hugging Face Daily Papers for August 6, from Shanghai Jiao Tong University: ABSeeker, Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment (arxiv.org/abs/2608.05102, 53 upvotes). The headline result: a 4B model trained on just 8.5k examples that scores 37.3 percent on BrowseComp and 39.1 percent on BrowseComp-ZH — 55.3 and 52.9 with context management — matching agents in the 30B class.
The problem it attacks is the oldest one in agent RL: a long search trajectory gets a single pass/fail reward at the end, so a 40-step run where 35 steps were smart and 5 were dumb gets judged as one undifferentiated failure. ABSeeker's answer, Answer-Backtracked Credit assignment, works backwards: start from the answer, trace back which intermediate clues were actually required to reach it, then score every search step against those clues. Sparse trajectory-level outcomes become dense step-level supervision — useful steps get rewarded even inside failed trajectories, redundant and wrong ones get suppressed.
Working backwards from the answer is one of those ideas that feels obvious the moment you hear it, which is usually the mark of a good one. The training data tells the story: 8.5k examples is nothing by RL standards, which means the win comes almost entirely from signal quality, not data volume. A 4B model punching at 30B is another data point for the year's recurring theme — the gap between model scale and agent performance is mostly a training-signal problem, and whoever assigns credit better gets to run smaller, cheaper agents. For anyone deploying search agents at scale, that arithmetic is the whole business case.
← Back to all articles
The problem it attacks is the oldest one in agent RL: a long search trajectory gets a single pass/fail reward at the end, so a 40-step run where 35 steps were smart and 5 were dumb gets judged as one undifferentiated failure. ABSeeker's answer, Answer-Backtracked Credit assignment, works backwards: start from the answer, trace back which intermediate clues were actually required to reach it, then score every search step against those clues. Sparse trajectory-level outcomes become dense step-level supervision — useful steps get rewarded even inside failed trajectories, redundant and wrong ones get suppressed.
Working backwards from the answer is one of those ideas that feels obvious the moment you hear it, which is usually the mark of a good one. The training data tells the story: 8.5k examples is nothing by RL standards, which means the win comes almost entirely from signal quality, not data volume. A 4B model punching at 30B is another data point for the year's recurring theme — the gap between model scale and agent performance is mostly a training-signal problem, and whoever assigns credit better gets to run smaller, cheaper agents. For anyone deploying search agents at scale, that arithmetic is the whole business case.
Comments