The most important command in BAAI's new research model is not search. It is update_context.

That command asks the model to stop accumulating pages for a moment and rewrite the state of the investigation: which findings have evidence, which candidates have been rejected, which constraints remain open, and what should be searched next. In AREX, verification is not reserved for the end of a report. It becomes the bridge from one research round to the next.[1]

The Beijing Academy of Artificial Intelligence released the AREX paper in late July 2026, then announced the research agent and its public platform on August 11. The package includes a 4-billion-parameter dense model called AREX-Turbo and an AREX-Base mixture-of-experts model with 122 billion total parameters and 10 billion activated per token. The Base and Turbo weights are available under Apache 2.0.[1][2][3][4]

As of August 30, 2026, the release is interesting for two separate reasons. The paper provides unusually direct ablations for the claim that a maintained research state improves web investigation. The open package also exposes a sharp boundary: BAAI has released a model trained to choose research actions, not a complete agent service that executes them for you.

Self-improvement here does not mean changing the weights

AREX begins with an asymmetry. Finding an answer that satisfies several entangled conditions can require a long search through weak clues. Once a candidate exists, checking each condition may be much easier. The model therefore runs an inner research loop to search, visit pages, assemble evidence, and produce a provisional answer. An outer self-improvement loop then accepts that answer, refines the useful trajectory around unresolved conditions, or discards a noisy trajectory and restarts.[1]

The phrase self-improvement can mislead. AREX is not rewriting its parameters while it researches. A Chinese launch account carried by BAAI Community is explicit: the model repeatedly revises its current answer and research state, not its weights.[3] A better operational description is recursive state repair.

The state repair has two levels. Inside a round, update_context replaces a swollen interaction history with a structured record of verified findings, citations, live and rejected candidates, validity concerns, unresolved constraints, and the next plan. Between rounds, the outer loop uses the provisional answer and a model-produced confidence score to decide whether to accept, refine, or restart. The paper caps evaluation episodes at 300 inner-loop turns and five outer-loop operations.[1]

This is more disciplined than simply granting an agent another hundred searches. Negative work survives: a candidate already disproved can remain in the refreshed state, reducing the chance that the model rediscovers and re-adopts it. Open questions survive too, so the next query can target a missing condition instead of vaguely asking for more information.

The ablation is more informative than the leaderboard

BAAI reports AREX-Base at 82.5% accuracy on BrowseComp, 89.9 F1 on DeepSearchQA, and 82.0 Item-F1 on the English WideSearch split. It also evaluates GAIA, xbench-2510, and Humanity's Last Exam with tools. Those are first-party results under BAAI's unified interface: search, visit, update_context, and finish, plus Python for HLE.[1]

The table is not a clean model-only race. Search providers, page parsers, prompts, tool budgets, and stopping rules are part of an agent result. The paper also marks some HLE competitors as evaluated on the full set while unmarked results use the text-only subset. “10B active parameters” describes routed compute in the MoE model; it does not make the 122B-weight checkpoint equivalent to a dense 10B model in storage, memory traffic, or deployment complexity.[1][2]

The controlled BrowseComp ablation says more about the release's actual contribution. With neither autonomous context updating nor the outer loop, the system scores 59.6. Adding context updating in the same single-round setting raises that to 71.4. Starting from that version, enabling the outer loop lifts accuracy again to 82.5. The complete system is therefore 22.9 points above the version with neither mechanism under the paper's matched setup.[1]

The runtime traces also support the claim that context updating is a research action, not just an emergency truncation. AREX invokes it in 80.3% of BrowseComp cases. The mean active context at an update is 25,721 tokens, far below the evaluation's configured 128,000-token ceiling. Search-strategy revision triggers 66.9% of updates, while candidate rejection triggers another 13.6%.[1]

There is an equally useful warning in the same analysis. Even with context updating, 33.0% of incorrect final outputs fall in the model's 90–100 confidence band. Confidence helps route many failures back into research, but it is not a calibrated certificate of truth. A production system should not let the model's own high score bypass source checks, independent validation, or human review for consequential work.[1]

BrowseComp measures persistence, not the whole job of research

BrowseComp contains 1,266 questions whose answers are short and verifiable but difficult to find through ordinary browsing. Its creators call the benchmark incomplete by design: it tests persistence and creative navigation while sidestepping parts of real user work such as ambiguity and long-form answers.[5] AREX's result is strong evidence about one demanding slice of research, not proof that the model can autonomously run a scientific program.

A newer literature-discovery benchmark makes the distinction concrete. AutoResearchBench asks agents either to trace a target paper through progressive clues or to retrieve an unknown set of papers meeting technical conditions. Its authors report that the strongest tested systems reached only 9.39% accuracy on the deep task and 9.31% intersection-over-union on the wide task.[6] AREX was not evaluated in that paper, so those numbers cannot be used as an AREX comparison. They show why performance on general web-search questions should not be silently extended to exhaustive literature review.

The launch coverage draws the boundary more clearly than the “autonomous research” label. The current AREX platform organizes news, papers, and podcasts and lets a user launch further investigation from that context. Proposal design, code execution, experiments, and performance tuning are described as future extensions.[3][4] Today, the public evidence supports an information-research agent. It does not yet establish an autonomous scientist that designs, runs, and reproduces experiments.

The weights are open; the machine around them is not included

The model card gives AREX-Base a 262,144-token context window and provides the weights, the complete BrowseComp prompts, and a minimal inference example. Its serving note uses eight-way tensor parallelism as a starting point with vLLM, SGLang, or another OpenAI-compatible server that supports Qwen3.5.[2] That is a substantial release: an operator can inspect the prompt contract and host the model without depending on BAAI's consumer application.

But the inference guide deliberately stops after generating the model's next action. When AREX emits an XML tool call, the caller must execute the real tool, append the assistant action and tool result to the conversation, and continue until finish appears.[2] The package does not supply a search provider, browser fetcher, page cleaner, citation resolver, credential boundary, retry policy, or production loop.

That missing layer is not plumbing in the dismissive sense. It determines which web the model can see and which evidence it is allowed to trust. It also decides whether a hostile page can inject instructions into the trajectory, whether a failed fetch becomes negative evidence, how duplicate sources are collapsed, how much time and money a round may consume, and when a confidence score is allowed to stop the run. Two teams serving the same checkpoint can build agents with meaningfully different reliability.

The release should therefore be read as an open research policy with a reference protocol. BAAI has made the decision-making model inspectable. Reproducing the paper's agent means reconstructing the environment around that model and matching its tools, budgets, prompts, and round logic closely enough that the learned behavior still has the conditions it expects.

A useful trial should test the state, not just the final answer

The cleanest internal evaluation would hold the search provider, page-visit implementation, token budget, and tool-call ceiling constant, then compare three versions: a single uninterrupted trajectory, a trajectory with update_context, and the full outer refine-or-restart loop. That mirrors the paper's central ablation instead of importing its score into a different stack.[1]

For each case, keep the intermediate research record. Check whether cited evidence survives compression, whether a rejected candidate reappears without new evidence, whether each unresolved constraint becomes a targeted action, and whether finish is called while material conditions remain open. Track tool calls, wall time, tokens, and failed visits alongside answer quality. A system that gains a few correct answers by spending five times the search budget has learned a different trade-off from one that redirects effort efficiently.

The most important falsifier is straightforward. If context updating and outer refinement do not improve constraint coverage or accuracy under a matched local budget—or if compression routinely drops decisive provenance—then AREX's strongest published mechanism does not transfer to that deployment. The right response would be to repair the tool and state contract, not to quote the 82.5 headline more loudly.

AREX's release delta is real and narrower than its name. It turns verification from a verdict into a routing instruction: keep what is proved, isolate what is missing, and make the missing condition the next problem. The open checkpoint can learn that rhythm. The operator still has to build the research room in which it can work.

Sources

1. Shuqi Lu et al., “AREX: Towards a Recursively Self-Improving Agent for Deep Research,” arXiv:2607.21461v2 (July 24, 2026) — architecture, training, evaluation protocol, benchmark results, context-update traces, confidence analysis, and ablations.
2. Beijing Academy of Artificial Intelligence, BAAI/AREX-Base repository (inspected August 30, 2026) — Apache 2.0 weights, architecture and context metadata, model family, complete BrowseComp prompts, and the instructions that leave tool execution to the caller: model card; minimal inference guide
3. 新智元, “智源AREX自主研究智能体：从「盲目搜索」到「验证驱动」” (August 2026; republished by BAAI Community) — Chinese launch coverage, benchmark framing, platform scope, and clarification that recursive improvement changes research state rather than model parameters.
4. Beijing International Science and Technology Innovation Center, “在关键时刻做对决定，北京智源研究院发布自主研究智能体” (August 12, 2026) — launch date, two-loop explanation, open-weight release, and current platform functions.
5. Jason Wei et al., “BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents,” arXiv:2504.12516 (April 16, 2025) — 1,266-question design, short-answer verification, and the benchmark authors' stated limits.
6. Lei Xiong et al., “AutoResearchBench: Benchmarking AI Agents on Complex Scientific Literature Discovery,” arXiv:2604.25256 (April 28, 2026) — literature-discovery task design and reported accuracy/coverage boundary for evaluated agents.
7. Li Ye, “重磅！智源研究院发布2025十大AI技术趋势,” Beijing Time (January 8, 2025) — source page for the documentary photograph of BAAI's Chengfu Road headquarters used as the cover image.