BAAI's AREX-2: a 27B open-weights model that runs its own try-check-retry agent loop
Sam Witteveen · youtube · 2026-11-06
Sam Witteveen reviews AREX-2, a new 27B open-weights model from BAAI fine-tuned from Qwen3-27B, built to run the try-check-retry loop itself rather than leaving it to the harness.
Key topics covered:
- Reflection and long-horizon capabilities
- The three-step data pipeline
- Benchmark results
- Skills vs training trade-offs
- Whether this counts as RSI (self-improvement)
- Hands-on testing (Round Lab) and takeaways
Weights on Hugging Face (BAAI/AREX-2), paper arXiv:2609.38288, code on GitHub (VectorSpaceLab/AREX-2).
More from coding & agent
- Frontier3DBench: 145-task benchmark tests coding agents on building interactive 3D worlds — _akhaliq · 2026-11-06
- Amazon research: self-distillation pushes weight memory to 59.9%, beating retrieval — omarsar0 · 2026-11-06
- Liquid AI's decision recipe matches Jev; new release adds on-device, modalities, primitives — helloiamleonie · 2026-11-06
- Decision models: general-purpose classifier or LLM? LoRA fine-tune wins — helloiamleonie · 2026-11-06
- LLMs are trained to predict next tokens, not make decisions — helloiamleonie · 2026-11-06
- Turn any LLM into a decision model with ~15 lines of inference-only code — helloiamleonie · 2026-11-06