HOT RADAR
llmHEAT 9.4°

AI CURATED HOT STORY · 10/27/2026

BAAI Releases AREX-2: 27B Model Achieves Cross-Domain Self-Improvement via Long-Horizon Reflection

1 reports archived1 independent sourcesupdated 10/27/2026, 12:03:00
Synthesis & Latest Updates
1 Sources Cross-Validated

BAAI and partners released AREX-2, built on Qwen3.8-27B. Trained with long-horizon reflection data, it sets new records for open-source lightweight models on benchmarks like MLE-bench and BrowseComp, demonstrating performance potential exceeding its parameter scale (compared to 1.6T models). The study proposes reframing Recursive Self-Improvement (RSI) as a trainable test-time scaling paradigm, validating long-horizon reflection as a transferable meta-capability.

LATEST/BAAI and partners released AREX-2, built on Qwen3.8-27B. Trained with long-horizon reflection data, it sets new records for open-source lightweight models on benchmarks like MLE-bench and BrowseComp, demonstrating performance potential exceeding its parameter scale (compared to 1.6T models). The study proposes reframing Recursive Self-Improvement (RSI) as a trainable test-time scaling paradigm, validating long-horizon reflection as a transferable meta-capability.

Hourly heat curve

11 fully observed hours
Now
9.4
Peak
11.713:00
24h change
–

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

Coverage timeline

Total 1 reports · Latest first
  1. 机器之心 (微信公众号)T2·85 pts
    • AREX-2 demonstrates that long-horizon reflection data enables cross-domain self-improvement, offering a verifiable training path for RSI.
    • The 27B model performs exceptionally well on multiple benchmarks, challenging the notion that parameter count dictates performance by showing high parameter efficiency.
    • Reframes RSI as Test-Time Scaling, emphasizing quality improvement in single tasks through continuous iterative reflection.