BAAI's new research agent doesn't just search longer, it checks its own homework

Deep research has an asymmetry that most agents ignore: finding the right answer is expensive, but verifying a candidate is cheap. BAAI, the Beijing Academy of Artificial Intelligence, built a family of agents that exploits this gap, the hidden costs of brute-force agent workflows make this contrast especially pointed. AREX, Recursively Self-Improving agents for deep research, searches smarter by treating its own intermediate outputs as work-in-progress that deserves a second look.

The architecture: a loop that inspects its own loop

AREX splits the research process into two nested cycles. The inner loop gathers evidence and assembles a provisional answer. The outer loop then audits that answer, checking each constraint that the original question imposes. When it finds an unresolved claim or a weak link, it launches targeted follow-up research, not another broad sweep, but a focused return to the specific gap.

Sustaining this over long horizons without losing track of what has been verified is the hard part. AREX solves it with a learned autonomous context-update tool that compresses the growing interaction history into a compact improvement state. Crucially, this compression does not rely on an external model; the agent learns to manage its own growth. The improvement state preserves verified evidence and unresolved constraints, letting the outer loop pick up exactly where it left off after each research iteration, addressing a known bottleneck in agent working memory.

Training for endurance, not just accuracy

BAAI trained AREX on verified synthetic tasks and high-quality trajectories using a two-stage recipe. First, agentic mid-training teaches the model to navigate multi-turn research environments. Then, long-horizon reinforcement learning shapes the agent to sustain improvement across many steps.

Reinforcement learning with sparse final rewards is notoriously hard for long sequences. AREX mitigates this by emphasizing key steps, moments where decisive evidence is acquired or where the agent corrects an erroneous research direction, a strategy that aligns with the orchestration-first approach recommended for small agentic models. This reward shaping gives the model a denser training signal without needing an external critic.

Punching above its weight

AREX comes in two sizes: a dense 4-billion-parameter model and a 122-billion-parameter Mixture-of-Experts version that activates only 10 billion parameters per token. On BrowseComp, WideSearch, DeepSearchQA, and Humanity's Last Exam (HLE), the paper reports that AREX substantially outperforms comparable-scale baselines and remains competitive with models using substantially more activated parameters. The 4B model demonstrates what efficient architecture and smart training can achieve without the massive compute budgets of proprietary labs, joining a growing list of compact models that challenge size-driven assumptions.

What this means for the deep research race

AREX arrives as the open-source momentum from East Asia accelerates. BAAI joins DeepSeek, Baichuan, and others in proving that frontier-capable research agents can be built without proprietary infrastructure. The recursive self-improvement loop, combined with learned context compression, offers a blueprint that other labs can adopt, the paper's code and model weights are available on the project page and Hugging Face. This trend has not gone unnoticed; open-source momentum from the East is challenging the US lead in AI.

The wider signal is that the gap between open and closed AI is narrowing not because proprietary models stagnate but because research agents are learning to get more out of less. AREX's 4B model challenges the assumption that deep research requires a 500-billion-parameter brain. If verification-guided refinement becomes a standard pattern, the next wave of research agents may be defined less by their size and more by how well they inspect their own work.