AI 代理與開放模型
AREX Uses Dual-Loop Verification to Drive Deep Research, Opens 4B and 122B-A10B Agent Models
The Beijing Academy of Artificial Intelligence has released AREX, a research agent that verifies tentative answers constraint by constraint, then launches another search round based on unmet requirements. The team also released Apache 2.0-licensed weights and runtime code, although the benchmark results still primarily rely on the authors’ own evaluations.
The Beijing Academy of Artificial Intelligence (BAAI) has unveiled the AREX family of deep-research agents. Rather than simply increasing the number of searches or extending context length, AREX divides research into alternating inner and outer loops. In the inner loop, an agent uses tools such as `search` and `visit` to gather information, synthesize evidence, and produce a tentative answer. The outer loop then checks each constraint in the original question and, based on confidence and identified gaps, decides whether to accept the answer, continue verifying specific claims, or discard an evidence-poor trajectory and restart.
To prevent long workflows from exhausting the context window, the model can also invoke `update_context` on its own, compressing verified findings, source identifiers, rejected candidate answers, unresolved constraints, and next steps into a new working state. Unlike fixed-length summarization, deciding what to retain is itself an action learned by the agent. The training pipeline combines agentic mid-training with long-horizon reinforcement learning and assigns greater reward weight to steps such as obtaining critical evidence and correcting an unproductive research direction. This is intended to mitigate the sparse feedback that results when rewards are assigned only to the final answer.
AREX-Turbo is a dense model based on Qwen3.5-4B, while AREX-Base uses the Qwen3.5-122B-A10B MoE architecture, activating approximately 10B parameters per token. Both are listed as supporting a 256K context window. Across six categories of evaluations—including BrowseComp, GAIA, DeepSearchQA, WideSearch, and HLE—the authors report that the models outperform several similarly sized baselines. BF16 weights, prompts, and quick-start scripts have also been released. Engineering teams should nevertheless interpret “recursive self-improvement” cautiously: the improvement here applies to an individual research trajectory and its compressed state, not to a deployed model modifying its own weights. The public repository currently has only a short history, and the search backend, tool reliability, inference costs, and whether benchmarks used equivalent budgets all require independent reproduction.