AREX-2: Advancing Self-Improving Agents through Long-Horizon Reflective Tasks
Hongjin Qian, Chaofan Li, Kun Luo, Wenqing Wei, Jianlyu Chen, Shuqi Lu, Yuyang Hu, Hongwang Xiao and 6 others
Beijing Academy of Artificial Intelligence (BAAI)
Published September 29, 2026 · on the feed October 1, 2026 · ▲ 128 · arXiv 2609.38288
The abstract
We present AREX-2, an effort to advance the self-improving capability of LLM agents, which we define as the ability to iteratively refine a solution at test time. This ability rests on two complementary capabilities: reflection, which produces a solution better than the current one, and long-horizon execution, which keeps the iteration effective over many rounds. We hypothesize that both capabilities are domain-agnostic, and can therefore be learned in scenarios that are well suited for supervision. Accordingly, we synthesize long-horizon improvement trajectories from machine learning and algorithmic programming tasks, two domains that offer verifiable feedback and reward sustained iteration. Trained on this data, our agent, built on Qwen3.8-27B, achieves strong results on MLE-bench Lite (81.8) and Frontier-CS (70.7), transfers to deep research with 84.0 on BrowseComp, 52.6 on HLE, 92.2 on GAIA, and 93.8 on DeepSearchQA, and keeps improving as its budget of rounds grows. These results show that long-horizon reflective data is an effective route toward self-improving agents.
The Brief
1:57A short podcast summary of the paper. Written and voiced by AI.
Transcript
AREX-2. Advancing Self-Improving Agents through Long-Horizon Reflective Tasks. By Hongjin Qian and thirteen others, from the Beijing Academy of Artificial Intelligence. Published on the twenty-ninth of September. Its code repository is worked out: about five hundred source files, with a readme and a licence.
The target here is a specific ability, which the authors define carefully: a language-model agent's capacity to refine a solution iteratively at test time. They call it self-improvement, and they split it into two capabilities that have to work together. Reflection produces a solution better than the current one. Long-horizon execution keeps that iteration effective over many rounds, rather than stalling after a few.
Their hypothesis is that both capabilities are domain-agnostic. If that is right, they need not be taught where they will eventually be used. They can be learned in whatever scenarios are best suited to supervision, and carried elsewhere.
The scenarios chosen are machine learning tasks and algorithmic programming tasks. Both offer verifiable feedback, and both reward sustained iteration. From them the authors synthesize long-horizon improvement trajectories, and train their agent on that data. The agent is built on Qwen3.8 with twenty-seven billion parameters.
On the home ground of the training data, the agent scores eighty-one point eight on MLE-bench Lite and seventy point seven on Frontier-CS.
The test of the hypothesis is transfer, and the agent carries over to deep research. It reaches eighty-four on BrowseComp, fifty-two point six on HLE, ninety-two point two on GAIA, and ninety-three point eight on DeepSearchQA. It also keeps improving as its budget of rounds grows.
The bet underneath AREX-2 is about where to teach a general skill. Reflection and persistence are learned on machine learning and programming, where every attempt can be checked, and then used on research tasks the training never covered. The results are offered as support for that bet: long-horizon reflective data as a route toward agents that improve themselves.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Hard problems are not solved in one attempt. Yet agents are trained as if they were.
You want an agent that, given more rounds on a hard task, keeps making its solution better by its own judgment. Today the harness decides when to retry; the model just makes one attempt at a time.
The road ahead
- 1 · The problem ~7 min Agents today make one attempt at a time while a harness decides when to retry, and training data throws away the work of improving.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Self-improvement needs reflection and long-horizon execution; both are domain-agnostic, so they are learned from whole improvement runs where feedback is checkable.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Pick checkable domains, synthesize long improvement runs there, train Qwen3.8-27B on them, then give it more rounds at test time.
- 6 · The experiment ~6 min You’ll guess: Trained on machine learning and programming improvement trajectories, what score did AREX-2 reach on the BrowseComp deep-research benchmark?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.