BAAI Introduces AREX, a Deep-Research Agent Designed to Check and Rerun Its Own Work

The Beijing Academy of Artificial Intelligence has introduced AREX, a family of deep-research agents designed to verify provisional answers and rerun research when gaps remain.

According to the arXiv paper, “AREX: Towards a Recursively Self-Improving Agent for Deep Research,” the system is built around a simple distinction: finding a plausible answer can be harder than checking whether a proposed answer satisfies each requirement in a question. AREX uses that distinction to structure research as an iterative process rather than a single pass.

A recursive research-and-verification loop

The AREX paper describes a two-part workflow. First, the agent performs web research, reads sources, gathers evidence, and drafts an answer. Then a verification step checks the answer against the task’s constraints, identifies unsupported or unresolved claims, and turns those gaps into new research objectives.

The VectorSpaceLab and BAAI project page describes AREX as “recursively self-improving” deep research agents in this sense: verified evidence is preserved, while unresolved constraints are used to guide the next round of research. The process can repeat until the agent has enough support for its answer or determines that a constraint cannot be resolved from available information.

This is different from simply retrieving more documents. The AREX paper frames the agent’s job as maintaining a loop between candidate answers, constraint-wise verification, and targeted follow-up investigation.

Managing long research sessions

A practical issue for recursive research systems is context growth. Long tasks can produce search results, partial notes, conflicting evidence, intermediate answers, and discarded hypotheses. The AREX paper says the system addresses this with an autonomous context-update tool that compresses interaction history into a more compact state.

The project page says this mechanism is intended to preserve verified findings and unresolved constraints across long-horizon tool use. In other words, AREX is designed to retain what has already been established while keeping track of what still needs investigation, rather than carrying every intermediate artifact forward unchanged.

Released models and availability

BAAI’s Hugging Face model card for AREX-Turbo describes it as a dense 4-billion-parameter deep-research agent. The card says the model is designed to search sources, assemble candidate answers, verify constraints, refine or restart research trajectories, and retain verified findings during long-horizon tool use.

The arXiv paper and Hugging Face Papers page also list AREX-Base, described in the candidate materials as a larger mixture-of-experts model with 122 billion total parameters and 10 billion activated parameters. Hugging Face Papers says model weights were released and that the models were integrated into the AREX online application.

Benchmark claims

The arXiv paper reports evaluations on web-research and reasoning benchmarks including BrowseComp, WideSearch, DeepSearchQA, Humanity’s Last Exam, reasoning tasks, and tool-use evaluations. The paper’s central claim is that recursive verification and targeted follow-up improve deep-research performance compared with comparable baselines across these evaluations.

Those results should be read as paper-reported benchmark findings rather than independent real-world validation. Still, the release is notable because BAAI is publishing both the research description and associated model artifacts, allowing other researchers and developers to inspect the approach more closely.

Why it matters

Many deep-research assistants already combine search, reading, and synthesis. AREX focuses on a narrower but important failure mode: a generated report may sound complete while leaving parts of the original question unverified. By turning unresolved constraints into the next research task, AREX attempts to make verification part of the agent’s operating cycle.

If the method proves reliable outside benchmark settings, it could influence how research agents are designed: less as one-shot answer generators, and more as systems that explicitly track evidence, uncertainty, and remaining work.