Hugging Face has released AREX (Automated Recursive Explorer), a new family of recursively self-improving (RSI) deep research agents designed to tackle complex multi-constraint problems. The work highlights a fundamental asymmetry in research: verifying a candidate answer is often easier than discovering it, suggesting agents should iteratively refine solutions rather than simply search longer.
AREX operates with two alternating loops: an inner research loop that gathers evidence and constructs a provisional answer, and an outer self-improvement loop that audits the answer constraint by constraint, identifies unresolved claims, and launches targeted follow-up research. To sustain long-horizon improvement, AREX learns an autonomous context-update tool that compresses interaction history into a compact state preserving verified evidence and unresolved constraints, without relying on an external model.
The models were trained on verified synthetic tasks and high-quality trajectories using agentic mid-training and long-horizon reinforcement learning. To mitigate sparse final rewards, the training emphasizes key steps where decisive evidence is acquired or erroneous directions are corrected. AREX comes in two sizes: a dense 4B parameter model and a 122B-A10B Mixture-of-Experts (MoE) model.
On benchmarks including BrowseComp, WideSearch, DeepSearchQA, and Humanity's Last Exam (HLE), AREX substantially outperforms comparable-scale baselines and remains competitive with models using significantly more activated parameters. The model weights are publicly available on Hugging Face, and an online application is live at arex-research.com.