27B AREX-2 Beats 1.6T Models: Long-Horizon Reflection Unlocks Trainable RSI Path

The Beijing Academy of Artificial Intelligence (BAAI), together with researchers from the University of Science and Technology of China and Renmin University, has released AREX-2, a 27B-parameter model built on Qwen3.8-27B. The model is trained using long-horizon reflection data constructed around machine learning and algorithmic programming scenarios. The core hypothesis is that recursive self-improvement (RSI) can be framed as a test-time scaling paradigm: letting an agent continuously reflect and iterate on its solution within a single task, turning more inference steps into better results.

RSI as Test-Time Scaling

The AREX team defines RSI not as a mysterious emergent property but as a measurable capability comprising two pillars:

Reflection quality – the ability to turn feedback into genuinely better proposals, not just self-critique.

Long-horizon persistence – the ability to sustain effective improvement over many rounds without stalling after failures or regressions.

These two factors act like the rate and time in compound interest, jointly determining the final return. The bottleneck is the scarcity of training data that fully demonstrates the continuous improvement process.

Machine Learning and Algorithmic Programming as Training Grounds

The team selected two domains that offer rich tasks and clear, verifiable environmental feedback:

Machine learning – agents improve models and training pipelines in GitHub repositories, validated by hidden evaluation metrics.

Algorithmic programming – agents optimize algorithm implementations on online judge problems, validated by test-case pass rates or solution quality.

A teacher agent first autonomously constructs tasks, scoring mechanisms, and runnable environments, filtering for those with sufficient headroom. Then, under generous time and step budgets, it runs loops of “submit solution → get feedback → reflect and improve,” optionally consulting references, using tools, and running experiments. The entire trajectory — reasoning, attempts, failures, and adjustments — is retained as training data that teaches “how to continuously improve.”

In-Domain Results: 27B Leads All Compared Models

On MLE-bench Lite , AREX-2 scores 81.8 , surpassing every model listed in the report, including closed-source frontiers. GPT-5.6 Sol and Kimi-K3 both score 72.7; AREX-2 leads by 9.1 percentage points. On Frontier-CS , AREX-2 achieves 70.7 , leading all reported open-weight models.

Cross-Domain Transfer: Deep Research Benefits Without New Domain Data

The critical test is whether the continuous improvement ability learned in ML and coding transfers to deep research. AREX-2 reuses AREX v1’s deep research data and adds newly synthesized long-horizon reflection trajectories from the two training domains — no new deep research data is added. On BrowseComp and HLE , AREX-2 scores 84.0 and 52.6 respectively, the highest among all models ≤40B parameters in the report.

This supports the team’s hypothesis that long-horizon reflection is a transferable meta-capability, not just a task-specific solving trick.

Parameter Efficiency: Smaller Model Reaches Higher Capability Tier

Plotting performance against parameter count reveals AREX-2’s exceptional parameter efficiency:

On MLE-bench Lite, 27B AREX-2 (81.8) beats 35B Frontis-MA1-35B (71.2) and 2.8T Kimi-K3 (72.7).

On BrowseComp, 27B AREX-2 (84.0) beats 1T Kimi-K2.6 (83.2) and 1.6T DeepSeek-V4-Pro (83.4).

The significance is not merely “small beats large.” It reveals an alternative capability growth path: besides scaling model size, dedicated training on the continuous improvement process lets models use feedback more effectively, correct solutions, and accumulate iterative gains.

Long-Horizon Reflection Data Provides a Verifiable RSI Training Path

AREX-2 demonstrates that long-horizon reflection data not only improves performance in the training domains but also helps models acquire a transferable self-improvement ability. For RSI, this offers a concrete, trainable, and verifiable direction: teach models to accumulate experience from every attempt and convert that experience into sustained progress over longer task horizons.

References

Paper: https://arxiv.org/abs/2609.38288 Project page: https://github.com/VectorSpaceLab/AREX-2 Model weights: https://huggingface.co/collections/BAAI/arex-2 Website:

https://arex-research.com/