BAAI's AREX-2 preprint trains long-horizon reflection on Qwen3.8-27B, scores 81.8 on MLE-bench Lite
xiaohu · x · 2026-11-10
A BAAI-led arXiv preprint, AREX-2, builds on Qwen3.8-27B to treat improvement-through-repeated-attempts as an agent training target, studying long-horizon reflection.
- Training trajectories contain multiple execution-feedback-revision iterations on ML and algorithmic coding tasks; earlier AREX deep-research data are retained, blending new reflective experience with existing task coverage.
- Reported scores: 81.8 on MLE-bench Lite, 70.7 on FrontierCS, 84.0 on BrowseComp, 52.6 on Humanity's Last Exam. These use different metrics and agent protocols, so they shouldn't be read as one accuracy scale or an equal-compute comparison.
- Self-improvement here includes revising solutions during test-time interaction; it does not imply recursive self-improvement of the model.
The figure outlines the pipeline: constructing environments from GitHub repos and online judges, building training data from multi-round trajectories, and test-time multi-round improvement after training.
More from Research
- Four Free Math Books for ML: Linear Algebra, Probability, Optimization, Information Theory — cneuralnetwork · 2026-11-10
- 4 free math books that actually matter for ML: linear algebra to information theory — cneuralnetwork · 2026-11-10
- Meta FAIR's e-means beats k-means in 1/10 epochs with 10-45x GPU speedups — burny_tech · 2026-11-10
- NeurIPS workshop paper: gradient-based data attribution can't reliably filter subliminal learning — ponguru · 2026-11-10
- Yang-Mills Working Group to Release OpenAI's Full Lean Code and Validation Tomorrow — burny_tech · 2026-11-10
- Interpretability researchers reminded: explanation matters, not just predictive power — burny_tech · 2026-11-10