ABSeeker: Dense Step-Level Credit From Backtracking the Answer Gets a 4B Search Agent to 55.3% on BrowseComp

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

arXiv 2608.05102 (submitted August 5, Lu, Ye, Wang et al.) converts sparse trajectory-level outcomes into dense per-step supervision by backtracking from the final answer to identify which actions actually helped, then rewards those and penalizes erroneous or redundant ones — implemented as both ABC-SFT (per-turn loss reweighting) and ABC-GRPO (step-level rewards in GRPO). With 8.5k training examples on Qwen3.5-4B it reports 37.3% on BrowseComp (55.3% with context management) and 39.1%/52.9% on BrowseComp-ZH, matching agents around 30B. It's the #1 paper on HuggingFace Daily Papers today at 47 upvotes, and the credit-assignment trick is the transferable idea for anyone training their own search or research agent.