BAAI's AREX-2 (27B) Sets Sub-40B SOTA on BrowseComp and HLE
A 27B model claiming Pareto-frontier results against trillion-parameter systems signals that targeted training on recursive self-improvement can substitute for raw scale. Teams running autonomous coding or deep-research agents can compare AREX-2's MLE-bench Lite and BrowseComp scores against their current backbone models before committing GPU budgets. Built on Qwen3.8-27B, it is also a candidate for local deployment where frontier-model API costs are prohibitive.
BAAI, with USTC and RUC, released AREX-2, a 27B model built on Qwen3.8-27B. It targets recursive self-improvement (RSI) as a test-time scaling paradigm, focusing on Deep Reflection and Long-Horizon Execution trained on algorithmic optimization tasks. AREX-2 scored 81.8 on MLE-bench Lite, 84.0 on BrowseComp and 52.6 on HLE, the best results among models at or under 40B parameters. The release positions a 27B model against 1T-2.8T parameter giants on autonomous research benchmarks.
Can a 27B model push the Pareto frontier of autonomous research and beat 1T–2.8T giants?
Meet AREX-2 (27B), spearheaded by BAAI in collaboration with USTC and RUC (built on Qwen3.8-27B). Turning Recursive Self-Improvement (RSI) into a practical test-time scaling paradigm, we pinpoint two keys to self-evolution: Deep Reflection & Long-Horizon Execution. By reinforcing algorithmic optimization tasks, AREX-2 emerges with exceptional deep reflection and long-horizon execution capabilities—significantly boosting its performance oncomplex coding and deep research to top MLE-bench Lite (81.8) and set new ≤40B SOTA on BrowseComp (84.0) & HLE (52.6)👇
