Shanghai Jiao Tong's 4B ABSeeker Rivals 30B Search Agents

Training a web-search agent is one of those problems where the signal you have and the signal you need are wildly mismatched. The agent takes many sequential steps to search, retrieve, verify and integrate evidence, and at the end you learn one bit: did it get the answer or not. Existing training pipelines smear that single bit uniformly across every step of the trajectory, so a successful run credits its redundant detours and a failed run punishes the steps that actually retrieved decisive evidence.

A group at Shanghai Jiao Tong University has posted a paper on Hugging Face proposing an unusually clean fix. Their method, Answer-Backtracked Credit Assignment (ABC), exploits a property of search tasks that other RL setups don't have: once you have the verified answer, the task is backtrackable. Their clue-recovery stage runs its own ReAct loop from the answer toward the query, using the same tool protocol as the forward agent, and pins down the intermediate entities, facts, and relations a valid trajectory should hit. Those clues become fixed anchors, and each step of a rollout is then scored on whether it discovered, verified, refined, or incorrectly reasoned about them. That converts sparse binary outcome supervision into dense per-step rewards, which they plug into both SFT (as loss reweighting, ABC-SFT) and GRPO (as step-level rewards, ABC-GRPO).

The numbers are the interesting part. Their model, ABSeeker, is trained on Qwen3.5-4B with only 8.5k examples. The paper reports 37.3% on BrowseComp and 39.1% on BrowseComp-ZH, and with context management enabled those climb to 55.3% and 52.9%. It also reports 77.0% on xbench-2505, 46.0% on xbench-2510, and 81.6% on GAIA-text. The authors' claim is that this beats other 4B baselines such as QUEST-4B and Dr.Venus and matches ~30B search agents including Tongyi DeepResearch and OpenSeeker.

The honest caveat is that these are self-reported benchmark numbers from a preprint, and "matches ~30B" is the authors' own reading of their bar chart, not an independent evaluation. The clue-recovery step also depends on an LLM to reconstruct evidence from the answer, so its quality on domains where the ground truth is genuinely obscure or contested is not something the paper's headline results address. What the reporting doesn't give you is training cost, inference-time compute, or robustness when the anchor answer itself is uncertain.

If the result holds up in independent hands, the real read is that step-level credit assignment, not scale, is where a lot of the remaining headroom for open search agents lives. That is good news for anyone trying to run capable deep-research agents on hardware that fits a single GPU.