研究代理／開放模型
AREX Enables Research Agents to Repeatedly Verify Unresolved Constraints, While a 4B Model Uses Compressed State for Long-Horizon Search
BAAI reframes deep research as a recursive loop of “search—verify each constraint—conduct targeted follow-up searches,” while training the model to compress confirmed evidence and unresolved constraints on its own. The weights for a 4B dense model and a 122B-A10B MoE model have been released, but their advantages mainly come from automatically verifiable synthetic tasks and cannot yet be directly generalized to open-ended scientific research.
The Beijing Academy of Artificial Intelligence (BAAI) and collaborators have released AREX, an effort to address deep-research agents’ tendency to lose track of constraints, duplicate work, or submit answers prematurely after prolonged searches. Rather than merely extending a single agent trajectory, the system alternates between two nested loops: an inner loop gathers evidence and generates a provisional answer, while an outer loop audits each constraint in the task, flags claims that remain unverified or contradict one another, and initiates targeted searches.
The biggest engineering obstacle in long-running workflows is continual context growth. AREX therefore learns a `context-update` tool that rewrites interaction history into a compact “improvement state,” preserving verified evidence, relationships among sources, and unresolved constraints instead of repeatedly feeding the complete browsing history back into the prompt. This compressor is learned by the agent itself, eliminating the need to call a larger external model during every iteration. In theory, this can reduce token costs and avoid state-translation errors between different models.
The training pipeline begins with synthetic research tasks that can be programmatically verified, followed by agentic mid-training and long-horizon reinforcement learning. To address excessively sparse rewards for final answers, the authors increase the learning weight of critical steps such as “obtaining decisive evidence” and “correcting an erroneous search direction.” The team released a dense version with approximately 4B parameters, as well as a Mixture-of-Experts (MoE) version with roughly 122B total parameters and about 10B parameters activated per token. The models were compared with similarly sized models on search, reasoning, and tool-use benchmarks including BrowseComp, WideSearch, DeepSearchQA, and HLE.
The technical significance lies in incorporating both memory management and verification strategies for research agents directly into model training, rather than simply scaling compute through longer contexts or more parallel agents. However, the paper’s term “recursive self-improvement” refers to iterative refinement of an answer within the same task; it does not mean the model can modify its own weights or improve its capabilities indefinitely. The evaluations also emphasize questions whose answers can be clearly verified. Assessing source quality, reconciling conflicting literature, accessing paywalled databases, and developing genuinely novel hypotheses remain outside their scope. Future work should examine whether compressed states silently discard counterevidence and whether external teams can reproduce the results under a fixed search budget.